1 Van Jacobson mentioned that hidden hops cause pathchar errors in his MSRI talk in 1997 [2].
We conclude in VIII. II. TAXONOMY
OF BANDWIDTH ESTIMATION TOOLS
In this section, we review the related work in the broader area of bandwidth estimation. Bandwidth estimation tools can be classied in several categories, depending on the specic throughput (or bandwidth) metric of interest, and on whether the measurements are performed on a per-hop or end-to-end basis. Table I summarizes all publicly available tools that we are aware of, together with their measurement objective and methodology. First, there are tools that measure end-to-end capacity. This is the maximum throughput that a path can provide to a ow when there is no other trafc in the path. The capacity is also referred to as bottleneck bandwidth. The underlying measurement methodology in such tools is usually a variation of packet pair or packet train probing. These methods are studied in detail in [10], [11], [12], [13], [14]. A second class is the tools that measure per-hop capacity. This is the maximum throughput that a single L3 hop can provide to a ow when there is no other trafc in that hop. The end-to-end capacity of a path is the minimum per-hop capacity, among all hops in that path. The underlying measurement methodology in such tools is the Variable Packet Size (VPS) probing technique, that we review in III. The available bandwidth of a network path is the maximum throughput that a path can provide to a ow, given the current trafc load in the path [15]. Measuring available bandwidth is much harder than measuring capacity, since the former is a dynamically varying metric. Measurement methodologies that attempt to estimate some form of available bandwidth have been proposed in [13], [15], [16], [17], [18], [19]. Another related throughput metric is the Bulk-TransferCapacity (BTC) [20]. The BTC of a path in a certain time period is the throughput of a bulk TCP transfer, when the transfer is only limited by the network resources and not by limitations at the end-systems. The BTC can be measured with Treno or cap [21]. Similarly, the throughput of large TCP transfers using parallel streams can be measured using Iperf [22], or similar TCP-based tools. III. VARIABLE PACKET S IZE (VPS)

PROBING

device of the path, which is the receiving host, is detected using the ICMP Port Unreachable option. Let us analyze the components of an RTT measurement (see Figure 1). Consider a path from a sender SN D to a receiver RCV that consists of H 1 hops. The capacity of each hop is Ci , for i = 1. H. The RTT from SN D to hop I for a packet of size L can be measured setting the TTL of the packet to I. The RTT TI (L) of the packet is

TI (L) =

L LICM P + if + df + + ir + dr i i r Ci Ci
In the previous equation, the fraction L/Ci is the transmission or serialization latency of a packet of size L at hop i. The terms if and df are the propagation and queueing delays, i respectively, of the packet at hop i of the forward path, from r SN D to hop I. The terms LICM P /Ci , ir , and dr are the i serialization, propagation, and queueing delays, respectively, of the ICMP reply at hop i of the reverse path, from hop I to SN D. For simplicity of notation, we assume that the forward and reverse paths go through the same sequence of links and routers; this assumption is not required by the VPS methodology, however. The serialization delays are introduced in store-and-forward packet forwarding devices, such as routers, because it takes L/Ci time units to transmit a packet of size L onto a link of capacity Ci ; we will return to this crucial point in V. The delay terms if and ir are due to the nite speed of propagation in physical-layer links, and they also include all constant per-packet processing in routers. The queueing delays df and dr can be introduced i i in router/switch buffers, when it is not possible to transmit a packet immediately. The VPS technique makes the following crucial assumption: if we measure the RTT TI (L) up to hop I with several probing packets, it is likely that the minimum RTT measurement TI (L) resulted from a packet, and a corresponding ICMP reply, which did not experience any queueing delays. When this is the case for a certain RTT measurement TI (L), we have that df = dr = 0, and so i i

TI (L) =

L LICM P + i + r Ci Ci
In this section, we briey review the VPS measurement methodology; a more comprehensive study is given in [4]. An important requirement in VPS probing is to be able to measure the RTT of a packet from the sender up to a certain hop I. This is possible using the Time-To-Live (TTL) eld of the IP header. Each L3 device along the path decrements the TTL before forwarding the packet to the next hop. If an L3 device receives a packet with zero TTL, it discards the packet and sends an ICMP Time Exceeded reply to the sender [7]. This technique is used by traceroute to identify the sequence of L3 devices in a network path [8]. VPS tools estimate the RTT up to each hop of the path sending packets with increasingly larger TTLs, and measuring the time interval until the receipt of the corresponding ICMP replies. We note that the last L3
where i = if + ir includes the constant delays at hop i, in both the forward and reverse paths. It is important that the ICMP replys size LICM P is constant, normally 32 bytes [7]2. Consequently, the minimum RTT measurement TI (L) can be expressed as I L (3) TI (L) = Di + Ci i=1 where Di = i + LICM P is independent of L. r Ci Suppose now that we measure the minimum RTT for different packet sizes L, with the constraint that L is not

2 The size of an ICMPv6 Time Exceeded packet depends on the size of the corresponding IPv6 probing packet, causing a problem for VPS probing.
Tool bprobe nettimer (A) sprobe pathrate pathchar clink pchar nettimer (B) pipechar cprobe pathload TReno cap IPerf
Author Carter [16] Lai [11] Saroiu [23] Dovrolis [14] Jacobson [2] Downey [24] Mah [5] Lai [6] Guojun [25] Carter [16] Jain & Dovrolis [19] Mathis [20] Allman [21] NLANR-DAST [22]
Measurement objective End-to-End Capacity End-to-End Capacity End-to-End Capacity End-to-End Capacity Per-Hop Capacity Per-Hop Capacity Per-Hop Capacity Per-Hop Capacity End-to-End Bottleneck End-to-End Avail-BW End-to-End Avail-BW Bulk-Transfer-Capacity Bulk-Transfer-Capacity Maximum TCP throughput
Methodology Packet Pairs Packet Pairs Packet Pairs Packet Pairs & Trains Variable Packet Size Variable Packet Size Variable Packet Size Variable Packet Size (tailgating) Packet Trains Packet Trains Self-Loading Periodic Streams Emulated TCP throughput Standardized TCP throughput Parallel TCP streams
TABLE I TAXONOMY OF PUBLICLY- AVAILABLE BANDWIDTH ESTIMATION TOOLS.

L/C1 C1

f d2 f 1

L/C2 C2

r L ICMP/ C1 r C1 r d1

L ICMP

I th Hop

Fig. 1.

RTT components in VPS probing.
larger than the path Maximum Transmission Unit (MTU)3. From Equation 3, it is easy to see that the relation between the minimum RTT and the packet size L is linear. Thus, TI (L) = I + I L (4) where I = Di , while I is the slope of the minimum RTT measurements TI (L) as L varies. The RTT slope at hop I is given by I 1 I = (5) Ci i=1 In VPS probing, we experimentally measure the RTT-slope I at each hop I of the path, and then calculate the perhop capacity CI from the RTT slope differences I I1 between successive hops. In more detail, suppose that we have measured the slope 1 at the rst hop. The capacity C1 can be then computed as C1 = 1/1. Applying induction, suppose that we have already estimated the capacity of the rst I hops using the slope measurements i , i = 1. I. The capacity of hop I + 1 can be then estimated as 1 (6) CI+1 = I+1 I from the RTT slopes I and I+1. This iterative approach can be applied, for example, in the RTT measurements of Figure 2 to estimate the capacities of a two-hop path.

IV. VPS

TOOLS AND VARIATIONS
In this section, we review four publicly-available per-hop capacity estimation tools, which are based on VPS probing. pathchar: The rst per-hop capacity estimation tool was pathchar, announced by V. Jacobson in 1997 [2]. The source code for pathchar was never released, however. It is known, though, that the tool follows closely the VPS methodology presented in the previous section. clink: This is an open source implementation of the VPS probing, presented by A. Downey in 1999 [24]. The primary differences between pathchar and clink are that the latter uses an even-odd technique, described in [4], to generate interval capacity estimates, and that when it encounters a routing instability, it collects data for all the paths it encounters, until one of the paths generates enough data to yield an estimate. pchar: This is also an open source implementation of the VPS probing methodology, announced by B. Mah in 1999 [5]. pchar runs on several Unix platforms, it provides kernellevel timestamps (via the pcap library), offers the option of three linear regression algorithms for the estimation of the RTT slopes, and it supports the use of several different types of probe packets.

fragmentation adversely affects the accuracy of VPS probing.

Higher Layers

Layer 3

Router

Layer 2 Physical Layer

Segment 1,1

Segment 2,1

Switch

Segment 2,2

Segment 2,3

Segment 3,1

Segment 3,2

Fig. 3.
The capacity of L3 hops and L2 segments in a network path.
First hop (Estimated capacity : 17Mbps) Second hop (Estimated capacity : 22Mbps)
the tailgating technique is as susceptible to layer-2 store-andforward devices as the previous three tools [26]. Also, the sources of error described in VII apply to the tailgating technique, but their exact effect is likely to differ from the analysis given there. V. T HE EFFECT OF L2 DEVICES

Minimum RTT (msec)

=0.84usec/byte 2

=0.47usec/byte 1

750 Packet size (bytes)

Fig. 2.

VPS probing example: RTT slopes at rst two hops.
nettimer: This tool is based on a combination of the VPS and packet pair probing methodologies [6]4. As in other VPS tools, nettimer sends probing packets of variable size, measures the minimum delay for each size, and estimates the RTT slope at a hop using linear regression. A major difference is that the RTTs do not require ICMP TimeExceeded replies from the path routers. Instead, nettimer uses an interesting packet tailgating technique, which only requires some cooperation from the receiving host. The tailgating code in nettimer is available, but its user-interface is not documented and not maintained. Consequently, the experimental results of VI do not include measurements with nettimer. In personal correspondence, the author of the tool has conrmed that
4 Note that nettimer is the name of a tool that can do either per-hop capacity estimation (using the packet tailgating technique of [6]), or endto-end capacity estimation (using the packet pair technique of [11]). Here, we focus on the former technique.
In this section, we show that L2 store-and-forward devices cause capacity underestimation in VPS probing. First though, it is important to distinguish between IP layer, or L3 links, and data-link layer, or L2 links. To differentiate between the two, we refer to L3 links as hops and to L2 links as segments. Each L3 hop includes at least one L2 segment, since the L3 connectivity is provided on top of L2 connectivity (see Figure 3). Different links are interconnected with packet forwarding devices, or simply devices. We characterize devices as L3 or L2, depending on whether they operate at the IP layer. Consequently, L3 devices decrement the Time-To-Live (TTL) eld of the IP header before forwarding the packet to the next hop, while L2 devices do not. So, L2 devices cannot be detected using ICMP, and they are essentially invisible to VPS probing. We note that IP routers are by denition L3 devices. There are also several L3 switch products which operate at the IP layer, providing the forwarding performance that was traditionally feasible only with L2 switches. Another important distinction is between store-and-forward devices and cut-through devices [27]. A store-and-forward device receives and buffers the entire packet before forwarding it to the next link. It is exactly this behavior that creates the serialization latency terms L/Ci in Equation 1. A cut-through device, on the other hand, needs to only receive a packets header before it starts forwarding the packet to the next link. In other words, cut-through devices overlap the receipt and the transmission of a packet, avoiding most of the serialization latencies. L3 routers and many L2 switches are store-andforward devices. Most hubs and physical-layer repeaters are cut-through devices, and, as will become clear later in this section, they have no impact on the accuracy of VPS probing. Our model of a network path, including both L3 and L2 devices, is shown in Figure 3. An L3 hop i consists, in general, of Mi 1 L2 segments. The rst segment of a hop is the outgoing interface in the corresponding L2 device. Additional L2 devices, such as switches of various technologies, can follow in that hop, however, before the next L3 device. The presence of such intermediate L2 devices (and segments) is

L1 /C a

L3 Ci =

j=1.Mi

L2 min {Ci,j }
Given that L2 devices are invisible to IP and higher-layer protocols, it is clear that VPS probing cannot measure the capacity of each segment, at least with IP and ICMP packets. L3 Instead, the objective is to estimate Ci for each hop i = 1. H of an end-to-end path. Let us now examine the impact of L2 store-and-forward devices on VPS probing. Suppose that the rst hop of the path consists of M1 segments, or equivalently, M1 store-andforward L2 devices. Each segment j = 1. M1 in that hop L2 introduces a latency of L/C1,j to a packet of size L, because of the serialization latency that store-and-forward devices cause. Consequently, following the derivations that led to Equation 3, the minimum RTT for a packet of size L is expected to be T1 (L) = D1 +

M1 j=1

Minimum RTT

L L2 C1,j

Fig. 4. An L2 device introduces an additional store-and-forward delay, causing capacity underestimation.
(8) RTT slope, as the lower part of Figure 4 shows, leading to capacity underestimation. Let us now examine whether L2 devices at a hop i can affect the capacity estimate of hop i + 1. This is an important issue, because the VPS methodology estimates the capacity of each hop using the RTT slope at the previous hop (see Equation 6). Suppose that the capacity of the rst hop has been underestimated as in Equation 10, and that the second L3 L2 hop consists of only one L2 segment with C2 = C2,1. The RTT slope at the second hop will then be 2 =

1 + L2 L2 C2,1 j=1 C1,j

Thus, the RTT slope of the rst hop is

1 L2 C1,j

and so the capacity estimate for the rst hop, according to VPS, is 1 L3 (10) C 1 = M1 1

L2 j=1 C1,j

L2 L3 as opposed to C1 = minj=1.M1 C1,j , which is the correct value of the capacity. It is easy to see that if M1 > 1 then L3 L3 C1 < C1 , meaning that the capacity of the rst hop will be underestimated. For the special case that all L2 segments L2 L2 have the same capacity (C1,j = C1 for j = 1. M1 ), we have that L2 L3 C (11) C1 = 1 M1

and from Equation 6 the capacity estimate for the second hop will be 1 L2 L3 = C2,1 (13) C2 = 2 1 which is the correct value. Even though the previous derivations were made for the rst two hops in a path, it is straightforward to apply them inductively to every hop in the path. To summarize the results of this section, the presence of more than one (Mi > 1) L2 store-and-forward devices at hop i causes capacity underestimation at hop i, but it does not affect the capacity estimate at hop i + 1, or at any subsequent hop.
Figure 4 illustrates the previous result with an example of one hop with two segments. The timeline shows the delays that a packet of size L1 would encounter in the rst hop, together with the delays of the corresponding ICMP reply. In this example, we also assume that the switch introduces a xed delay ; that constant delay, however, has no impact on the RTT slope and the capacity estimate. Instead, it is the switchs serialization latency that causes the increased

common in LANs and campus networks. L2 switches are also used in some backbone networks to implement a fullyconnected topology at the IP layer. L2 Each segment j of hop i has a capacity of Ci,j bits-persecond (bps), meaning that it can transmit a packet of size L L2 bits in L/Ci,j seconds. In the rest of the paper we assume constant link capacities, ignoring technologies such as certain wireless links or trafc shapers, in which the capacity can vary with the underlying error rate or trafc burstiness. The L3 capacity Ci of hop i is dened as the minimum of the L2 segment capacities in that hop.

Layer 2 Device

L1 /C b

Segment 1,1 C1,1=C a

Segment 1,2

C1,2 =C b >C a

2=1/C a+1/C b
With intermediate L2 device
Additional delay due to L2 device

L1 /Cb

1 =1/ Ca + L Without intermediate L2 device

L1 /Ca

Probing packet size

L1/L c L c
In this section, we present experimental data that resulted from the publicly-available VPS tools pathchar, clink, and pchar on local-area, campus-wide, and wide-area network paths. A rst objective in these experiments is to verify the negative effect of L2 store-and-forward devices on the accuracy of VPS probing. A second objective is to examine whether the problem of L2 devices appears only in local-area networks, which are often built exclusively based on Ethernet switches, or whether the problem is more general. A third objective is to observe the nature and magnitude of errors that are not related to L2 devices; such errors are investigated in more detail in VII. In the following, we report the capacity estimates that resulted from 15-30 independent runs of each tool, as ranges. A special case of cut-through devices is that of ATM switches. In ATM, an IP packet is segmented into a number of xed-size cells, after some ATM Adaptation Layer (AAL) encapsulation. The segmentation takes place at the last store-andforward device in the path, before entering the cloud of ATM segments. An ATM segment transfers cells independently, and so it does not wait to receive all cells of a packet before forwarding them to the next segment. Suppose that Lc is the ATM cell size, and L1 is the probing packet size, ignoring for simplicity the extra bytes for AAL encapsulation. The timeline for the transmission of the packet through a hop that includes an ATM switch is shown in Figure 5. The timeline looks like a staircase, because a packet of size L1 is fragmented into L1 /Lc ATM cells. What is most important though, is that the ATM switch increases the packets RTT by a constant, without affecting the overall RTT slope on which the VPS capacity estimate is based on. So, ATM switches do not affect the accuracy of VPS probing, as long as the probing packet size can be much larger than the ATM cell size Lc. Cut-through L2 devices forward a packet to its next segment after receiving a xed-length initial part of the packet. That part includes, normally, only the L2 (or MAC) header, which always appears at the start of a packet. Let us denote by LH the initial number of bytes that a cut-through device needs to receive, before forwarding a packet of size L (LH < L). Also, let C L2 be the capacity of the corresponding segment. Note that the segment will introduce a delay LH /C L2 in the processing of a packet, as opposed to L/C L2 that a storeand-forward would introduce. This is because cut-through devices overlap the receipt and transmission of a packet, after the initial LH bytes have been received and processed. Consequently, cut-though devices introduce a constant delay term in the RTT of a packet, so they do not affect the RTT slope or the capacity estimate of VPS probing. B. ATM switches A. Cut-through L2 devices VI. E XPERIMENTAL

orion.pc.cis 100Mbps 100Mbps

128.175.137.66

100Mbps

128.4.132.64

chpbr4g500.nss

chpbr4f101.nss

Host Incoming Interface FastEthernet Switch

chprt1v29.nss

10Mbps

chprt1v9.nss

Ethernet Switch
4) The capacity estimates that we can calculate from Equation 10, for hops with known L2 segments are close to the measurements of VPS tools, even though the agreement is far from perfect in some cases. As an illustration of the last point, consider the path from orion to tsunami. The rst hop has M1 =5 Fast Ethernet segments. According to Equation 10, the VPS capacity estimate is expected to be 100/5=20Mbps, which is close to the 17Mbps estimate of the VPS tools. Admittedly, though, the VPS measurements are closer to what one would expect with 6 Fast Ethernet segments (100/616.7Mbps). Similarly, the capacity of the second hop in the same path is measured in the 55-75Mbps range, while Equation 10 predicts only 50Mbps. The third hop is a Fast Ethernet interface which is estimated correctly by all tools, as there are no intermediate L2 devices; the measurements have a signicant variation however (20Mbps). For the fourth hop we would expect 5Mbps, while the tools measure around 5.6Mbps. The measurements for the last hop indicate the presence of at least one switch, but we were unable to verify that part of the L2 infrastructure. In the reverse path, from tsunami to orion, we do not have complete knowledge of the L2 infrastructure in the rst two hops. The results seem to indicate though the presence of an Ethernet switch in the rst hop, and the absence of any switches in the second hop. The third hop in the path from tsunami to orion, which is a Gigabit Ethernet link, gives widely varying results, even though it does not include L2 segments. Some other possible error sources are discussed in VII. At the fourth hop of that path we expected 50Mbps, but the measurements varied around 40Mbps. Finally, the presence of four intermediate Fast Ethernet switches at the last hop should lead to a capacity estimate of 20Mbps. It is only pchar that measured something close to that value, however. C. Access and Wide Area Network (WAN) links We also attempted to measure per-hop capacities in several WAN paths. Unfortunately, we could not get consistent capacity estimates using any of the VPS tools in such paths. The results often vary by two orders of magnitude, from a few Mbps to hundreds of Mbps, indicating strong probabilistic sources of measurement errors. The nature of such errors is the subject of the next section. For now, we simply present capacity measurements for the two access links of the UnivDelaware, as well as for two links in the corresponding backbone providers (Abilene and VoiceNet). The rst entry at Table V refers to the Univ-Delaware access link to VoiceNet. The outbound network interface at the Univ-Delaware gateway is a Fast Ethernet (chp-br4-f1-0-1.nss.udel.edu), which is rate-limited to 45Mbps at the VoiceNet end. Note that the VPS tools underestimate that hops capacity at about 30Mbps. This value can be explained if we assume that there is an intermediate Fast Ethernet switch in that hop at the VoiceNet end of the hop, because 1/45 + 1/100 1/31 (see Equation 10). Unfortunately, we could not verify with the VoiceNet engineers whether such a Fast Ethernet switch exists at their end of that hop. The second

newarkgw

chp7ke24.nss

10Mbps tsunami.coastal

Fig. 6. The L3 and L2 topologies for two paths at the Univ-Delaware campus. All names share the udel.edu sufx. The dashed arrows represent hops for which we do not know the underlying L2 infrastructure.
The capacity estimates of the VPS tools are given in Tables III and IV. We summarize those results with the following observations: 1) The three VPS tools are usually consistent with each other. 2) The VPS tools are accurate in measuring Ethernet and Fast Ethernet hops, when there are no intermediate L2 switches in those hops. 3) The capacity is signicantly underestimated in all hops that include intermediate L2 switches.
5 As a side-note, packet dispersion techniques can also be signicantly errorprone along heavily-loaded paths [14].
L3 hop from orion.ps.cis to 128.4.132.64 from 128.4.132.64 to chp-br4-f-1-0-1.nss from chp-br4-f-1-0-1.nss to chp-rt1-v-9.nss from chp-rt1-v-9.nss to chp-7k-e-2-4.nss from chp-7k-e-2-4.nss to tsunami.coastal
Nominal capacity 100Mbps 100Mbps 100Mbps 10Mbps 10Mbps

TABLE III

PATH FROM
pathchar 17.00.0 62.27.2 100.515.0 5.750.15 4.50.1
clink 17.00.0 64.79.3 100.322.0 5.60.1 3.70.1
pchar 17.00.4 62.39.1 101.926.0 5.70.1 6.50.6
C APACITY ESTIMATES FOR THE

orion.pc.cis.udel.edu

tsunami.coastal.udel.edu.
L3 hop from tsunami.coastal to newark-gw from newark-gw to chp-rt1-v-29.nss from chp-rt1-v-29.nss to chp-br4-g-5-0-0.nss from chp-br4-g-5-0-0.nss to 128.175.137.66 from 128.175.137.66 to orion.pc.cis
Nominal capacity 10Mbps 10Mbps 1000Mbps 100Mbps 100Mbps
pathchar 4.050.05 10.50.5 613.33150.0 38.31.7 6.950.5
clink 4.00.0 10.80.4 414.70580.0 39.96.0 6.10.2
pchar 4.01.2 11.10.9 450.2110.0 35.68.8 21.57.8
TABLE IV C APACITY ESTIMATES FOR THE tsunami.coastal.udel.edu orion.pc.cis.udel.edu.
L3 hop from chp-br4-f-1-0-1.nss.udel.edu to delaware-gw-f2-0.voicenet.net from delaware-gw-f2-0.voicenet.net to delaware2-gw-H2-0-T3.voicenet.net
Nominal capacity 45Mbps 45Mbps

TABLE V

pathchar 30.53.5 44.620.0

clink 30.35.6 48.01.6

pchar 28.35.6 45.210.0
C APACITY ESTIMATES FOR THE U NIV-D ELAWARE ACCESS LINK TO V OICE N ET, AND FOR A V OICE N ET EDGE LINK.
hop at Table V is an edge T3 link in the VoiceNet network. All three VPS tools manage to estimate its capacity correctly, but with signicant statistical variation in the case of pathchar and pchar. Table VI refers to the PoS OC-3 access link from UnivDelaware to Abilene, and to a PoS OC-48 core link in the Abilene network. The VPS tools consistently underestimate the capacity of the OC-3 link to a value that is close to 80Mbps. Unfortunately, we could not get information about the exact L2 infrastructure at the Abilene end of the hop. We note, however, that the presence of an intermediate OC-3 L2 switch in the measured hop would result (from Equation 10) to a capacity estimate of approximately 155/2 = 77.5Mbps, which is close to what the three tools measure. Finally, the VPS tools do not manage to get a reasonable estimate for the

OC-48 core link in the Abilene network. The capacity of that link is very high (2.48Gbps), and as shown in the next section, the ability of VPS tools to measure links in that bandwidth range is limited by the measurement hosts clock resolution.

VII. OTHER

SOURCES OF ERROR
In this section, we examine some other sources of error in VPS probing. As will become clear next, these errors are probabilistic, in the sense that successive measurements can lead to widely different results, as opposed to the consistent underestimation errors caused by L2 devices. In the following, we assume that there are no intermediate L2 devices in the measured path.
L3 hop from chp-br4-f-1-0-1.nss.udel.edu to abilene-wash-gsr.nss.udel.edu from abilene-wash-gsr.nss.udel.edu to atla-wash.abilene.ucaid.edu
Nominal capacity 155Mbps 2480Mbps

pathchar 82.200

clink 82.63.410

pchar 82.67.800

TABLE VI C APACITY ESTIMATES FOR THE U NIV-D ELAWARE ACCESS LINK TO A BILENE , AND

FOR AN

A BILENE OC-48 CORE LINK.

Path length I 8 10

=0.380

=0.21959

=0.22486182 P (I, K) 0.9.
is likely that some of the reported errors in VPS tools are due to non-zero queueing delays, and they may be avoided with a larger number of probing packets, especially for links that are further away along the path7. B. Effect of non-zero queuing delays As shown in the previous paragraph, high utilization or long paths can introduce non-zero queuing delays, even in the minimum RTT measurement. Suppose that we want to estimate the capacity C of a link with only two probing packet sizes L1 and L2. Let the minimum RTT measurement at these packet sizes be T1 = + L 1 + q1 C q2 T2 = + L 2 + C (16) (17)
TABLE VII M INIMUM NUMBER OF PACKETS K

SO THAT

A. Effect of trafc load As mentioned in III, a key assumption in VPS probing is that the minimum RTT measurement for each packet size does not include any queueing delays. If the network load is signicant, however, this may not be true even with a large number of probing packets and RTT measurements. To examine this issue quantitatively, let us assume that probing packets arrive at each link as a Poisson process6. Even though this is a crude assumption, our objective here is simply to illustrate that measuring a queueing-free RTT can be quite hard in loaded paths. Based on the Poisson Arrivals See Time Averages (PASTA) property [28], the probability that a probing packet will not experience any queueing at a link i is 1 i , where i is the fraction of time that the link i is busy, or equivalently, the utilization of link i. Consequently, the probability that a probing packet will not experience any queueing delay at the rst I links of the path is

where q1 and q2 are the minimum queue sizes seen by probing packets of size L1 and L2 , respectively. The RTT slope that we would measure then, is T T2 T 1 q = =+ L L2 L 1 CL (18)
where = 1/C, T = T2 T1 , L = L2 L1 , and q = q2 q1. So, the estimated capacity will be C L = C= q T 1 + L (19)

P (I) =

(1 i )
and so, the probability that at least one out of K probing packets will not see any queuing delay in the rst I links is P (I, K) = 1 [1 P (I)]
Table VII shows the number of probing packets K per link and per packet size, so that the probability that at least one probing packet sees no queuing delays is more than 90%. The utilization i is assumed to be the same at all links. Note that K is impractically large when the path includes more than 5-6 links that are 60%, or more, loaded. We note that the default number of probing packets per link and per packet size is 32 for pathchar, 8 for clink, and 32 for pchar. Consequently, it
6 This may not be true even when probing packets are sent from their source as a Poisson stream, because the interarrivals of probing packets can be modied in the network.
Therefore, non-zero queueing delays cause a multiplicative error term in the capacity estimate. This error factor can be reduced with a larger packet size variation L; usually however, L is limited by the 1500 byte Ethernet MTU. It is important to note that all VPS tools measure the RTT slope at a link using tens of different packet sizes and elaborate linear regression algorithms. Consequently, VPS tools are more robust to queueing delays than what the previous model implies. C. Effect of limited clock resolution The accuracy of RTT measurements is also limited by the resolution of the clock at the measurement host. If the clock resolution is 2, any time instant between t0 and t0 + will be measured as t0. The minimum RTT measurement for two
7 The transmission of probing packets should be of course rate-limited to avoid self-queueing.
packet sizes L1 and L2 can then be anywhere in the following ranges T1 = + L 1 T2 = + L 2 In the worst-case, the RTT slope can be estimated as 2 T = L L The estimated capacity would then be C= C 1 2C L (23) (22) (20) (21)
E. Effect of the ICMP messages generation latency It is a conventional wisdom that VPS tools are sometimes inaccurate because router ICMP replies, which are required for the RTT measurements, are generated from slow processing and forwarding paths. First of all, this may not even be the case, given the recent measurements of [30]. Even if it is true for some routers, however, it should be clear from III that it is not the latency of generating ICMP replies that can affect the RTT slope measurements. Instead, it is the variation of those latencies that can affect the measured RTT slope. In other words, ICMP messages will not affect the accuracy of VPS probing as long as the minimum ICMP reply generation latency TICM P at a router does not vary with the size L of the packet that caused the ICMP reply. If this is the case, the impact of ICMP packet generation on VPS probing is no different than that of queueing delays: a sufciently large number of probing packets is required, so that the minimum RTT measurement corresponds to an ICMP reply with the minimum generation latency TICM P. VIII. R ELATED WORK

AND CONCLUSIONS

Table VIII shows the range of capacity estimates that can result at high-bandwidth links, when the clock resolution is 2=1sec, and the packet size variation is limited by the Ethernet MTU (L 1500 bytes). Such a high resolution is typical for workstations today8. Note that with these values of and L, it is basically futile to accurately measure OC-48 or higher bandwidth links with a VPS tool. D. Error propagation from previous links In VPS, the capacity estimate for a link i depends on the RTT slope that was measured at link i 1 (see Equation 6). To examine how estimation errors can propagate along a path, consider a two-hop path with C1 and C2 being the capacities of the two links. Ideally, the RTT slopes should be measured as 1 = 1/C1 at the rst link, and 2 = 1/C1 + 1/C2 at the rst two links. Suppose now that the rst link introduces an error, bounded by , in the RTT slope measurements. Then, even if there is no error introduced at the second link, the two RTT slopes would be measured as 1 = 1 (1 +

1 ), 2

= 1 (1 +
where 1 and 2 are the errors introduced by the rst link (| 1 |, | 2 | ). The capacity of the second link will be estimated as C2 = C= 1+( 2

C) C1 2

The estimation error is maximized when C2 = C+ 2 C2 /C1

=. Then, (26)

The previous expression shows that if the capacity ratio C2 /C1 of two successive links is larger than one, any estimation error at the rst link is magnied by that capacity ratio at the second link. For example, if an Ethernet link is followed by a Gigabit Ethernet link (C2 /C1 =100), an error factor of 0.1% at the rst link can result in a 10% error in the capacity estimate of the second link.
8 A resolution of just a few microseconds is often based on interpolation between successive clock interrupts, meaning that it may not be exact [29].
This paper examined different sources of error in VPS probing. We focused on the effect of L2 store-and-forward devices, which introduce serialization latencies just like L3 routers, but without decrementing the TTL eld of probing packets. It has been shown that such L2 devices introduce signicant and consistent estimation errors, which are impossible to detect unless the L2 path infrastructure is known. Some other sources of error, such as queueing delays, limited clock resolution, propagation of errors from previous hops, have been also examined. An important point is that these latter effects are probabilistic. So, when the effects of these error sources are large enough, repeated runs of a VPS tool over the same path will produce signicantly different capacity estimates. In this sense, probabilistic error sources in capacity estimation are easier to detect. A different technique to measure the capacity of each L3 hop would be to send pairs (or trains) of probing packets with an expiring TTL, and then measure the resulting dispersion at each hop. This technique, which is similar to what pipechar does [25], is not affected by the presence of L2 devices in the path. The problem with this technique, however, is that it cannot measure the capacity C of a link if there is a previous link in the path with capacity C < C. Recently, [31] presented ve new methodologies for estimating per-hop capacities, using sequences of four packets called packet quartets. A packet quartet consists of two independent packet-pairs, with the rst packet of each pair, called pacesetter, followed by a much smaller packet called probe. The pacesetters TTL is set to expire at a particular hop in the path. The proposed methodologies are based on the end-to-end delay variation of the two probe packets. Two of those methods are referred to as PQ1 and PQ2. In both PQ1 and PQ2, the pacesetters expire at the same hop. In PQ1 the probes have the same size, while in PQ2 the pacesetters have the same size. The probe packet delay variations in PQ1 are up

[11] K. Lai and M. Baker, Measuring Bandwidth, in Proceedings of IEEE INFOCOM, Apr. 1999, pp. 235245. [12] V. Paxson, End-to-End Internet Packet Dynamics, IEEE/ACM Transaction on Networking, vol. 7, no. 3, pp. 277292, June 1999. [13] B. Melander, M. Bjorkman, and P. Gunningberg, A New End-to-End Probing and Analysis Method for Estimating Bandwidth Bottlenecks, in Global Internet Symposium, 2000. [14] C. Dovrolis, P. Ramanathan, and D. Moore, What do Packet Dispersion Techniques Measure? in Proceedings of IEEE INFOCOM, Apr. 2001, pp. 905914. [15] M. Jain and C. Dovrolis, End-to-end available bandwidth: measurement methodology, dynamics, and relation with TCP throughput, in Proceedings of ACM SIGCOMM, Aug. 2002. [16] R. L. Carter and M. E. Crovella, Measuring Bottleneck Link Speed in Packet-Switched Networks, Performance Evaluation, vol. 27,28, pp. 297318, 1996. [17] V. Ribeiro, M. Coates, R. Riedi, S. Sarvotham, B. Hendricks, and R. Baraniuk, Multifractal Cross-Trafc Estimation, in Proceedings of ITC Specialist Seminar on IP Trafc Measurement, Modeling, and Management, Sept. 2000. [18] G. Jin, G. Yang, B. Crowley, and D. Agarwal, Network Characterization Service (NCS), in Proceedings of 10th IEEE Symposium on High Performance Distributed Computing, Aug. 2001. [19] M. Jain and C. Dovrolis, Pathload: A measurement tool for endto-end available bandwidth, in Proceedings of Passive and Active Measurements (PAM) Workshop, Mar. 2002, pp. 1425. [20] M. Mathis and M. Allman, A Framework for Dening Empirical Bulk Transfer Capacity Metrics, July 2001, RFC 3148. [21] M. Allman, Measuring End-to-End Bulk Transfer Capacity, in Proceedings of ACM SIGCOMM Internet Measurement Workshop, Nov. 2001, pp. 139144. [22] NLANR Distributed Applications Support Team, Iperf Version 1.2, http://dast.nlanr.net/Projects/Iperf/, May 2001. [23] S. Saroiu, Sprobe: A Fast Tool for Measuring Bottleneck Bandwidth in Uncooperative Environments, http://www.cs.washington.edu/homes/tzoompy/sprobe/, Sept. 2001. [24] A. Downey, clink: a Tool for Estimating Internet Link Characteristics, http://rocky.wellesley.edu/downey/clink/, June 1999. [25] J. Guojun, Network Characterization Service, http://wwwdidc.lbl.gov/pipechar/, July 2001. [26] K. Lai, Effect of L2 Devices on Nettimer, Personal Correspondence, July 2002. [27] J. Kurose and K. Ross, Computer Networking: a Top-Down Approach Featuring the Internet. Addison-Wesley, 2001. [28] R. Wolff, Poisson Arrivals See Time Averages, Operations Research, vol. 30, no. 2, pp. 223231, 1982. [29] A. P sztor and D. Veitch, A Precision Infrastructure for Active a Probing, in Proceedings of Passive and Active Measurements (PAM) workshop, 2001. [30] R. Govindan and V. Paxson, Estimating Router ICMP Generation Delays, in Proceedings of Passive and Active Measurements (PAM) workshop, Mar. 2002, pp. 613. [31] A. P stzor and D. Veitch, Active probing using packet quartets, in a Proceedings of Internet Measrement Workshop (IMW), Nov. 2002, pp. 293306. [32] R. Govindan and H. Tangmunarunkit, Heuristics for Internet Map Discovery, in Proceedings of IEEE INFOCOM, Mar. 2000, pp. 1371 1380. [33] S. Savage, A. Collins, E. Hoffman, J. Snell, and T. Anderson, The End-to-End Effects of Internet Path Selection, in Proceedings of ACM SIGCOMM, Sept. 1999, pp. 289300.

5U, 19" rack mounted mixer
" R A C K M I X E R M I T 5 H E MEZCLADOR DE 5U, 19" FORMATO RACK MIXER RACKABLE 19'' X 5U
OPERATIONS MANUAL BEDIENUNGSHANDBUCH MANUAL DEL OPERADOR MANUEL DINSTRUCTIONS
MULTI LANGUAGE INSTRUCTIONS
ENGLISH........PAGE 6 DEUTSCH........PAGE 9 ESPAOL..........PAGE12 FRANCAIS...........PAGE 15
PLEASE READ BEFORE USING APPLIANCE, IMPORTANT WARNING & SAFETY INSTRUCTIONS!

CAUTION

RISK OF ELECTRICAL SHOCK DO NOT OPEN!
CAUTION: This product satisfies FCC regulations when shielded cables and connectors are used to connect the unit to other equipment. To prevent electromagnetic interference with electric appliances such as radios and televisions, use shielded cables and connectors for connections. The exclamation point within an equilateral triangle is intended to alert the user to the presence of important operating and maintenance (servicing) instructions in the literature accompanying the appliance. The lightening flash with arrowhead symbol, within an equilateral triangle, is intended to alert the user to the presence of uninsulated dangerous voltage within the products enclosure that may be of sufficient magnitude to constitute a risk of electric shock to persons. READ INSTRUCTIONS: All the safety and operating instructions should be read before the product is operated. RETAIN INSTRUCTIONS: The safety and operating instructions should be retained for future reference. HEED WARNINGS: All warnings on the product and in the operating instructions should be adhered to. FOLLOW INSTRUCTIONS: All operating and use instructions should be followed. CLEANING: The product should be cleaned only with a polishing cloth or a soft dry cloth. Never clean with furniture wax, benzine, insecticides or other volatile liquids since they may corrode the cabinet. ATTACHMENTS: Do not use attachments not recommended by the product manuacturer as they may cause hazards. WATER AND MOISTURE: Do not use this product near water, for example, near a bathtub, wash bowl, kitchen sink, or laundry tub; in a wet basement; or near a swimming pool; and the like. ACCESSORIES: Do not place this product on an unstable cart, stand, tripod, bracket, or table. The product may fall, causing serious injury to a child or adult, and serious damage to the product. Use only with a cart, stand, tripod, bracket, or table recommended by the manufacturer, or sold with the product. Any mounting of the product should follow the manufacturers instructions, and should use a mounting accessory recommended by the manufacturer. CART: A product and cart combination should be moved with care. Quick stops, excessive force, and uneven surfaces may cause the product and cart combination to overturn. See Figure A. VENTILATION: Slots and openings in the cabinet are provided for ventilation and to ensure reliable operation of the product and to protect it from overheating, and these openings must not be blocked or covered. The openings should never be blocked by placing the product on a bed, sofa, rug, or other similar surface. This product should not be placed in a built-in installation such as a bookcase or rack unless proper ventilation is provided or the manufacturers instructions have been adhered to. POWER SOURCES: This product should be operated only from the type of power source indicated on the marking label. If you are not sure of the type of power supply to your home, consult your product dealer or local power company. LOCATION: The appliance should be installed in a stable location. NON-USE PERIODS: The power cord of the appliance should be unplugged from the outlet when left unused for a long period of time. GROUNDING OR POLARIZATION: If this product is equipped with a polarized alternating current line plug (a plug having one blade wider than the other), it will fit into the outlet only one way. This is a safety feature. If you are unable to insert the plug fully into the outlet, try reversing the plug. If the plug should still fail to fit, contact your electrician to replace your obsolete outlet. Do not defeat the safety purpose of the polarized plug. If this product is equipped with a three-wire grounding type plug, a plug having a third (grounding) pin, it will only fit into a grounding type power outlet. This is a safety feature. If you are unable to insert the plug into the outlet, contact your electrician to replace your obsolete outlet. Do not defeat the safety purpose of the grounding type plug. POWER-CORD PROTECTION: Power-supply cords should be routed so that they are not likely to be walked on or pinched by items placed upon or against them, paying particular attention to cords at plugs, convenience receptacles, and the point where they exit from the product. OUTDOOR ANTENNA GROUNDING: If an outside antenna or cable system is connected to the product, be sure the antenna or cable system is grounded so as to provide some protection against voltage surges and built-up static charges. Article 810 of the National Electrical Code, ANSI/NFPA 70, provides information with

regard to proper grounding of the mast and supporting structure, grounding of the lead-in wire to an antenna discharge unit, size of grounding conductors, location of antenna-discharge unit, connection to grounding electrodes, and requirements for the grounding electrode. See Figure B. LIGHTENING: For added protection for this product during a lightening storm, or when it is left unattended and unused for long periods of time, unplug it from the wall outlet and disconnect the antenna or cable system. This will prevent damage to the product due to lightening and power-line surges. POWER LINES: An outside antenna system should not be located in the vicinity of overhead power lines or other electric light or power circuits, or where it can fall into such power lines or circuits. When installing an outside antenna system, extreme care should be taken to keep from touching such power lines or circuits as contact with them might be fatal. OVERLOADING: Do not overload wall outlets, extension cords, or integral convenience receptacles as this can result in a risk of fire or electric shock. OBJECT AND LIQUID ENTRY: Never push objects of any kind into this product through openings as they may touch dangerous voltage points or short-out parts that could result in a fire or electric shock. Never spill liquid of any kind on the product. SERVICING: Do not attempt to service this product yourself as opening or removing covers may expose you to dangerous voltage or other hazards. Refer all servicing to qualified service personnel. DAMAGE REQUIRING SERVICE: Unplug this product from the wall outlet and refer servicing to qualified service personnel under the following conditions: When the power-supply cord or plug is damaged. If liquid has been spilled, or objects have fallen into the product. If the product has been exposed to rain or water. If the product does not operate normally by following the operating instructions. Adjust only those controls that are covered by the operating instructions as an improper adjustment of other controls may result in damage and will often require extensive work by a qualified technician to restore the product to its normal operation. If the product has been dropped or damaged in any way. When the product exhibits a distinct change in performance, this indicates a need for service. REPLACEMENT PARTS: When replacement parts are required, be sure the service technician has used replacement parts specified by the manufacturer or have the same characteristics as the original part. Unauthorized substitutions may result in fire, electric shock, or other hazards. SAFETY CHECK: Upon completion of any service or repairs to this product, ask the service technician to perform safety checks to determine that the product is in proper operating condition. WALL OR CEILING MOUNTING: The product should not be mounted to a wall or ceiling. HEAT: The product should be situated away from heat sources such as radiators, heat registers, stoves, or other products (including amplifiers) that produce heat.

PDM-01 0

PDM-02 0

INTRODUCTION:

Congratulations on purchasing a Gemini PDM series 19"5U, 4 channel, rack mounted audio mixer. This state of the art mixer is backed by a 3 year warranty, excluding the cross fader. The cross fader is backed by a separate 90 day warranty. Prior to use, we suggest that you carefully read all the instructions.
- Alternatively, the BALANCED MASTER (6) output jacks also connects the mixer to the main amplifier using standard cables with 1/4" TRS connectors. We recommend using balanced cables if the distance to your amp is 10 feet or more. - The BOOTH (7) output jacks allow the connection of an additional amplifier with RCA cables. - The ZONE (5) output jacks allow the connection of an additional amplifier with RCA cables. - The REC (4) output jacks can be used to connect the mixer to the record input of your recording unit, thus enabling you to record your mix. 4. Located on the rear panel are 2 PHONO (PH) /LINE (LN) convertible RCA inputs (12, 16), & 5 LINE RCA INPUTS (10, 9, 13, 17, 18). The convertible RCA inputs for CH 2 (16) & CH 3 (12) allow PH and LN level equipment to be connected to the mixer. To adjust the CONVERTER(s) (11, 15), just flip the switch UP to operate PH 1 or PH 2. Flip the switch DOWN to operate through LN 2 or LN 4. The PH INPUTS only accept turntables with a magnetic cartridge. When using (a) turntable(s), you will need to ground the RCA cable(s) by screwing in the grounding fork(s) to the GROUNDING SCREW (14) located in the back panel of the PDM mixers. This is located in between the CONVERTER SWITCHES (11, 15). The stereo LN INPUTS only accept line level inputs such as a CD, DAT, MiniDisc, etc.
NOTE: WHEN USING TURNTABLES, NOT ATTACHING A GROUND MAY CAUSE A SYSTEM "HUM."

FEATURES:

- 5U, 19" rack mounted mixer - 4 stereo channels - 7 lines, 3 Mic, 2 phono/line convertible RCA inputs - Master, record, booth, & zone RCA outputs - 1/4" balanced master output - 2 x 1/4" Mic inputs - Dual 10 band graphic EQ with on/off switch & blue LED indicator - Rotary zone, booth, & cue volume controls - Stereo/mono switch - Assignable cross fader - Removable, user replaceable Rail Glide cross fader - Push button cue section per channel with green LED indicator - CUE/PGM fader control allowing cue mix - XLR-1/4" combo Mic input - 2 band rotary Mic EQ & volume controls - Talk over feature - BNC lamp port - 1/4" headphone jack
NOTE: ABOVE FEATURES INCLUDED IN EACH MODEL IN THE PDM SERIES.
5. Headphones may be plugged into the face-plate located 1/4" HEADPHONE JACK (26). 6. The MIC 1 (49) input (located on the face panel) is a combination XLR & 1/4 connector. The MIC 2 (20) & MIC 3 (19) inputs (in the rear panel) accept only 1/4" connectors. The mic inputs accept balanced & unbalanced connections. 7. The BNC LAMP PORT (22) (located on the face panel, above the POWER SWITCH (21)) is used to plug in a 12 V BNC goose neck lamp such as the Gemini GNL-700.

PDM-01 FACE:

- Master volume line fader control

PDM-02 FACE:

- 6 digital samples with volume, & speed rotary controls - Rotary master volume control - Master/Mic assignable echo effect switch with repeat & delay controls

PDM-03 FACE:

- 96 second digital sampler comprised of 5 memory banks with soft touch backlit sample buttons - Rotary assignable channel control for sample recording - Sample parameter control with rotary level, pitch bend, record/single/repeat controls - Push button RoboPlay & cue sampler - Rotary master volume control

OPERATING INSTRUCTIONS:

1. Once all of your connections have been made in the rear panel, turn ON the mixer by pressing the POWER SWITCH (21). Once turned ON, the POWER BAR LED, containing the power symbol located in the VU METER (23), will be illuminated. Turn OFF the mixer when not in use by pressing the POWER SWITCH (21) to OFF. When the PDM mixer is turned OFF the POWER BAR LED will not be illuminated. 2. CHANNEL (CH) 1: To bring this channel into program mix (PGM), you must first decide which LN will be in use. Use the LN SWITCH (37) to toggle from LN 1 (18) to MIC 3 (19) on this channel. Slowly raise the CH 1 SLIDE CONTROL (39) to a comfortable level, once you've selected the proper line. 3. CH 2: To bring this channel into PGM, you must first decide which LN will be in use. Use the LN SWITCH (40) to toggle from PH 1/LN 2 (16) to LN 3 (17) on this channel. Slowly raise the CH 2 SLIDE CONTROL (42) to a comfortable level, once you've selected the proper line. 4. CH 3: To bring this channel into PGM, you must first decide which LN will be in use. Use the LN SWITCH (43) to toggle from PN 2/LN 4 (12) to LN 5 (13) on this channel. Slowly raise the CH 3 SLIDE CONTROL (45) to a comfortable level, once you've selected the proper line. 5. CH 4: To bring this channel into PGM, you must first decide which LN will be in use. Use the LN SWITCH (46) to toggle from LN 6 (10) to LN 7 (9) on this channel. Slowly raise the CH 4 SLIDE CONTROL (48) to a comfortable level, once you've selected the proper line. 6. CUE: By connecting a set of headphones to the HEADPHONE (26) jack, you can monitor any or all channels. Press the CUE BUTTONS (38, 41, 44, 47) for CHs 1 through 4, respectively, to assign the CH(s) to be monitored. The respective CUE LED indicators will glow when in use. Use the rotary CUE VOLUME CONTROL (25) to adjust the CUE volume without changing the overall mix. By moving the CUE/PGM FADER CONTROL (24) to the LEFT you will be able to monitor the assigned CUE signal. Moving the CUE/PGM FADER CONTROL (24) to the MIDDLE allows CUE mix with PGM. Moving the CUE/PGM FADER CONTROL (24) to the RIGHT allows you to monitor PGM output. 7. ASSIGN: There are 2 rotary controlled X FADER ASSIGN SWITCHES (34, 35), each having 5 settings OFF, 1, 2, 3, & 4. The LEFT (34) not

PRECAUTIONS:

1. All operating instructions should be read before using this equipment. 2. To reduce the risk of electrical shock, do not open the unit. Please refer servicing to a qualified service technician.
IN THE USA ~ IF YOU EXPERIENCE PROBLEMS WITH THIS UNIT CALL GEMINI CUSTOMER SERVICE AT: 1 (732) 738-9003. DO NOT ATTEMPT TO RETURN THIS EQUIPMENT TO YOUR DEALER.
3. Do not expose this unit to direct sunlight or to a heat source such as a radiator or stove. 4. This unit should be cleaned only with a damp cloth. Avoid solvents or other cleaning detergents. 5. When moving this equipment, it should be placed in its original carton and packaging. This will reduce the risk of damage during transit. 6. DO NOT EXPOSE THIS UNIT TO RAIN OR MOISTURE. 7. DO NOT USE ANY SPRAY CLEANER OR LUBRICANT ON ANY CONTROLS OR SWITCHES.

CONNECTIONS:

1. Before plugging the power cord in, make sure that the VOLTAGE SELECTOR (1) switch is set to the correct voltage. 2. Located on the rear panel is the 115 V/230 V PLUG (2). Before plugging the power cord in, make sure the POWER SWITCH (21) located on the face panel is turned OFF. 3. The PDM mixers have 5 sets of outputs: - The MASTER (3) output jacks connects to the main amplifier with RCA cables.

PDM SERIES

ASSIGN switch allows you to direct CH 1, 2, 3, or 4 through the LEFT side of the CROSS FADER (36). The RIGHT ASSIGN (35) switch allows you to direct CH 1, 2, 3, or 4 through the RIGHT side of the CROSS FADER (36). When the ASSIGN SWITCH(es) (34, 35) are at OFF, you will not have a CH assigned to the CROSS FADER (36). This allows you to control the PGM with the use of the respective CH SLIDE CONTROLS, thus layering the PGM with up to four CHs. 8. CROSS FADER SECTION: The CROSS FADER (36) allows you to mix from one source to another. The PDM mixers feature an assignable CROSS FADER (36). The rotary controlled ASSIGN SWITCHES (34, 35) allow you to select which channel will play through each side of the CROSS FADER (36). The CROSS FADER (36) in your unit is removable & if the need arises can be easily replaced. Your Gemini mixer comes with a RG- 45 (RAILGLIDE) DUAL-RAIL CROSS FADER. RAIL GLIDE CROSS FADERS have internal dual stainless steel rails that allow the slider to ride smoothly and accurately from end to end. Also available is our RG-45 PRO (PROGLIDE) CROSS FADER with a special curve designed for scratch mixing. Just purchase one from your Gemini dealer & follow the instructions: also plug a second & third MIC into the rear panel's MIC 2 (20) & MIC 3 (19) 1/4" jacks. The decibel level of MIC 2 (20) is controlled by the rotary MIC 2 VOLUME CONTROL (51). The decibel level of MIC 3 (19) is controlled by the CH 1 SLIDE CONTROL (39). 13. TALKOVER: The purpose of the AUTO TALKOVER MODE is to allow the program playing to be attenuated so that the MIC may be heard above the music. The AUTO TALKOVER SWITCH (54) controls MIC 1 (49) and MIC 2 (20) with 3 settings: - When the MIC/TALKOVER SWITCH (54) is in the BOTTOM position, MIC 1 (49), MIC 2 (20) & TALKOVER MODE are all OFF. - When the MIC/TALKOVER SWITCH (54) is in the CENTER position, MIC 1 (49) & MIC 2 (20) are ON & TALKOVER MODE is OFF. The MIC ON LED indicator glows when MIC 1 (49) & MIC 2 (20) are ON. - When the MIC/TALKOVER SWITCH (54) is in the TOP position, MIC 1 (49) & MIC 2 (20) are ON, TALKOVER MODE is ON, & the volume of all sources except MIC 1 (49) & MIC 2 (20) are lowered automatically by -16 dB, when speaking into the MIC(s). 14. BNC LAMP PORT: The BNC LAMP PORT (22) connects a 12 V BNC goose neck lamp, such as the Gemini GNL-700 to the PDM mixer. The goose neck lamp will be powered by your mixer. To turn ON the goose neck lamp, you must first attach the goose neck lamp to the BNC LAMP PORT (22). Make sure the PDM mixer is OFF when connecting the 12 V BNC lamp. To connect the goose neck lamp, simply align the screw cap of the goose neck lamp to the locking nodules of the BNC LAMP PORT (22), push down, & twist the screw cap clockwise to lock the 12 V BNC goose neck lamp in place. Then turn ON your mixer. The goose neck lamp should light-up. To detach the goose neck lamp from the BNC LAMP PORT (22), first make sure your mixer is OFF. Turn OFF your mixer and the goose neck lamp will turn OFF. Unscrew the screw cap by twisting it counterclockwise, then pull up & remove the goose neck lamp. 15. GROUND LIFT SWITCH: The GROUND LIFT SWITCH (8) is used to reduce background noise & hum when using multiple outlets to power audio equipment. The switch should be in the position that provides the least amount of noise or hum. If noise remains at the same level in both positions, the GROUND LIFT SWITCH (8) should be kept in the GND position.

NOTE: MAKE SURE THE MIXER AND/OR AMPLIFIER IS OFF BEFORE SWITCHING THE GROUND LIFT SWITCH TO PREVENT A TRANSIENT POP THAT MAY DAMAGE YOUR SYSTEM.

REPLACEABLE CROSS FADER

1. UNSCREW THE OUTSIDE FADER PLATE SCREWS (B). - DO NOT TOUCH INSIDE SCREWS (C). 2. CAREFULLY REMOVE OLD CROSS FADER AND UNPLUG CABLE (D). 3. PLUG IN THE NEW CROSS FADER INTO CABLE (D) AND PLACE BACK INTO MIXER. 4. SCREW THE CROSS FADER TO MIXER WITH THE FADER PLATE SCREWS (B).
NOTE: DO NOT APPLY PRESSURE WHILE USING THE CROSSFADER. LIGHTLY GLIDE THE CROSSFADER BACK AND FORTH. PRESSING DOWN ON THE CONTROLS CAN BEND CONTACTS AND CAUSE A LOSS OF SOUND.
9. EQUALIZER (EQ): These units feature dual 10 BAND GRAPHIC EQUALIZERS (32, 33) that will allow you to adjust the sound to fit any room. By adjusting any of the 10 EQ SLIDE CONTROLS (32, 33), you can cut or boost the tonal characteristics of the sound coming from PGM to the speaker(s) by 12 dB. To activate the dual 10 BAND GRAPHIC EQ, switch the EQ SWITCH (31) to ON, & the EQ LED will light up to indicate that the EQ has been engaged. To deactivate the dual 10 band graphic EQ, switch the EQ SWITCH (31) to OFF, & the EQ LED will turn OFF. When activated, the EQ (32, 33) controls the LEFT and RIGHT side of your stereo speakers. The PGM & EQ are controlled by the MASTER VOLUME (27). To balance the sound of the PGM playing through the MASTER VOLUME (27) on the LEFT & RIGHT side of your speakers you must mirror the EQ levels on the LEFT (32) & RIGHT (33) EQ controls.
NOTE: FOR OPTIMAL PERFORMANCE IN YOUR SOUND OUTPUT, HAVE YOUR SOUND SET TO STEREO NOT MONO. START WITH THE EQ LEVELS (32, 33) AT CENTER VALUE. THE EQ SLIDE CONTROLS (32, 33) SHOULD LOCK AT THIS POSITION. ADJUST YOUR MASTER VOLUME (27) CONTROL FROM MID TO LOW VOLUME RANGE. THEN ADJUST THE LEFT (32) OR RIGHT (33) EQ, ONE SLIDE CONTROL AT A TIME, TO A COMFORT ABLE LEVEL. ONCE YOU ARE SATISFIED WITH THE SOUND OF ONE SIDE, MATCH THE EQ SETTINGS ON THE OTHER SIDE. ONCE YOU HAVE PASSED THE CENTER VALUE ON THE EQ (32, 33), THE MASTER OUTPUT, AS INDICATED IN THE VU METER (23), MAY EXPERIENCE A TONAL BOOST. PLEASE ADJUST THE MASTER VOLUME (27) TO A COMFORTABLE LEVEL SO YOU DO NOT OVERLOAD YOUR SYSTEM. CLIPPING WILL OCCUR WHEN YOU ARE OVERLOADING YOUR SYSTEM. LOWER THE MASTER VOLUME (27) OR ADJUST YOUR EQ (32, 33) SETTINGS SO THAT CLIPPING DOES NOT OCCUR. THEN YOU MAY RAISE THE MASTER VOLUME (27) TO A LEVEL WITH WHICH YOU ARE COMFORTABLE.

16. VU METER: The VU METER (23) indicates the decibel level of the MASTER RCA & MASTER BALANCED (3 & 6) outputs of the LEFT & RIGHT stereo levels.

PDM-02 ECHO/EFX:

ECHO SECTION:
An echo effect may be applied to the PGM, or MIC 1 (49) & MIC 2 (20) signals by switching the ECHO ASSIGN (55) switch from MIC 1-2 on the LEFT, to OFF in the MIDDLE, to MASTER on the RIGHT & vice versa. When using ECHO (55), you may adjust the effect of the ECHO (55) by using the rotary REPEAT (56), DELAY (57), and ECHO VOLUME (58) controls. To turn the ECHO ASSIGN (55) OFF or lower the ECHO VOLUME (58).

SOUND EFFECTS SECTION:

Six different sound effects (APPLAUSE, SCREAM, COPTER, SCRATCH, H2O & GLASS) may be added to your mix by pressing the SOUND EFFECTS CONTROL BUTTONS (61). The volume of the effects can be adjusted using the rotary EFX VOLUME (60) located above the APPLAUSE effect button. The pitch of the effects can be increased or decreased using the rotary SPEED CONTROL (59) located above the SCRATCH effect button.
10. STEREO/MONO: You can convert your sound output from STEREO to MONO & vice versa by using the STEREO/MONO SWITCH (30). Switch to the LEFT to convert to STEREO. Switch RIGHT to convert to MONO. 11. OUTPUT SELECTION CONTROL: Once you are comfortable with the sound level of your music you may adjust the decibel level of the PGM with the MASTER VOLUME (27) control. MASTER RCA & BALANCED MASTER OUTPUTS (3, 6) are controlled by the level of the MASTER VOLUME (27) control. You may adjust the volume of the ZONE (5) output with the ZONE (29) rotary control. You may adjust the volume of the BOOTH (7) output with the BOOTH (28) rotary control. The volume of your RECORD (4) output is controlled strictly by the CH SLIDE CONTROLS. 12. MIC SECTION: Plug your main MIC into the MIC 1 combination XLR-1/4" input (49) located on the face panel. The rotary controls for HIGH (52) and LOW (53) allow you to adjust the tone of MIC 1 (49). The rotary MIC 1 VOLUME CONTROL (50), above the rotary MIC 2 VOLUME CONTROL (51), adjusts the decibel level of MIC 1 (49). You may
PDM-03 SAMPLER OPERATION:

MEMORY INFORMATION:

The PDM-03 is equipped with 5 MEMORY BANKS (59). The two banks marked 8 & 8 are 8 seconds in length, the two banks marked 16 & 16 are 16 seconds in length and the bank marked 48 is 48 seconds in length. These banks are separate & CANNOT be linked. You can store a different sample in each bank, but they must be recorded individually & must be played one at a time.

SAMPLE RECORDING:

1. Select the sample source by switching to the appropriate channel with the rotary SAMPLER ASSIGN (55) control. 2. The PDM-03 comes equipped with a rotary sampler PITCH CONTROL (56). To get a "perfect" sample, set the control to its CENTER position & record the sample. Raising or lowering the control during playback raises or lowers the pitch of the sample. The CENTER position retains the "normal" pitch.

HINT: YOU CAN RECORD A SAMPLE WITH THE PITCH CONTROL IN ANY POSITION. THAT POSITION SETTING WILL BECOME THE NORMAL PITCH. IF YOU START TO RECORD A SAMPLE WITH THE PITCH CONTROL SET AT "MINIMUM" THAT WILL BECOME YOUR NORMAL PITCH. BY INCREASING THE PITCH TO "MAXIMUM" DURING PLAYBACK, THE PITCH EFFECT WILL DOUBLE IN SPEED. RECORDING AT "MAXIMUM" AND LOWERING TO "MINIMUM" DURING PLAYBACK WILL DO EXACTLY THE OPPOSITE.
must be stopped. Choose the MEMORY BANK to play. Then press SAMPLER to begin the new Sample playback. 5. Tapping the SAMPLER (59) button with the MODE SELECTOR (58) switch in the REPEAT position causes the SAMPLE to loop play continuously. In REPEAT mode, the SAMPLER button acts an ON/OFF switch. The first push starts the sample, the second push stops it. 6. The SAMPLER LEVEL (57) controls the decibel level of the sample. This feature allows you to adjust the volume of the sample to play over or under PGM.

ROBO PLAY:

1. With the ROBO (59) button OFF the blue LED is OFF & the MODE SELECTOR (58) switch in either the SINGLE or REPEAT mode, pressing the SAMPLER (59) button will cause the sample to play along with the selected source. 2. When the ROBO (59) button is ON, the BLUE LED is on & starting the sampler mutes the selected source. When the sample ends, the source automatically turns back on.
3. Put the MODE SELECTOR (58) switch into the RECORD position. 4. Listen to the channel to be recorded in CUE, by selecting the appropriate CUE button for this channel. When the track approaches the section to be sampled, press the proper MEMORY BANK (59) button where you want the sample to be stored. 5. Then press the SAMPLER (59) button to start recording the sample. The Memory bank in use will have a blinking LED, if a battery is not in place or the battery is low (See BATTERY BACKUP section). The sample will be stored in this MEMORY BANK (59) & ready to play.
NOTE: TAPPING THE SAMPLER BUTTON BEGINS THE SAMPLING PROCESS (THE SAMPLER INDICATOR WILL "GLOW" BLUE DURING RECORDING). TAPPING THE SAMPLER BUTTON A SECOND TIME ENDS THE SAMPLE (THE SAMPLER INDICATOR WILL TURN OFF). IF YOU DO NOT TAP THE SAMPLER BUTTON A SECOND TIME, THE SAMPLING PROCESS STOPS AUTOMATICALLY AFTER 8, 16 OR 48 SECONDS DEPENDING ON WHICH MEMORY BANK WAS SELECTED.

BATTERY BACKUP:

1. BATTERY BACKUP: The PDM-03 is equipped with battery backup to retain samples. To activate this feature, a 9 V battery (not included) should be connected to the BATTERY HOLDER (60) located on the rear panel. This enables the storage of samples in memory. When the unit is unplugged, the battery backup retains the samples for future use.

NOTE: IF THE UNIT IS UNPLUGGED WITH NO BATTERY ATTACHED, ALL SAMPLES WILL BE LOST.

CUE SAMPLER:

1. To record in CUE, press the CUE SAMPLER button and then press the CUE button for the channel to be sampled. Be sure your CH slide controls are at zero so the sample does not play in PGM. Follow the SAMPLE RECORDING instructions to complete the process. 2. To test the recorded sample before playing in PGM, press the CUE SAMPLER (59) button, placing the sampler in CUE. The blue CUE SAMPLER LED will illuminate while in use. Use the CUE controls to monitor this sample. Be sure your rotary SAMPLER LEVEL (57) is turned counter clockwise so that the sample is not played in PGM. Set your sampler to SINGLE or REPEAT, then press the SAMPLER button and the sample will begin playing in CUE. If you are satisfied with your sample, leave it stored in the MEMORY BANK (59). If not, please repeat the steps for SAMPLE RECORDING.
2. LOW BATTERY INDICATOR: When the selected memory bank LED blinks, this will indicate that there is a low battery or no battery in the PDM-03. The LED blinks a warning if no battery is connected to the PDM-03. When changing or placing the battery into the PDM-03, make sure the unit is plugged in and the power is ON. Failure to adhere to this will result in lost memory and "vanished" samples.

SPECIFICATIONS:

INPUTS:
Phono:...3 mV, 47 KOhm Line:...150 mV, 27 KOhm MIC 1, 2, & 3:...1.5 mV, 2 K Ohm Balanced Bass:.... 12dB High:... 12dB

SAMPLE PLAYBACK:

1. Set the MODE SELECTOR (58) switch to SINGLE or REPEAT. 2. Select the desired sample by pressing the proper MEMORY BANK (59) button. 3. Tapping the SAMPLER (59) button with the MODE SELECTOR (58) switch in the SINGLE position causes SAMPLER to play back one time (the SAMPLER INDICATOR will "glow" GREEN). Each push of the SAMPLER button restarts the sample from the beginning. Rapid pressing of the SAMPLER button will cause a stuttering effect. Once the sample has started playback & the SAMPLER button is not pushed a second time, the SAMPLE will SINGLE to the end & stop. This Sample will play through completely regardless of switching the MEMORY BANK. Switching the MEMORY BANK and pressing the SAMPLER button while the sample is in play will repeat the sample previously selected until it has completed its play cycle. 4. To play a new sample from another MEMORY BANK the old sample

OUTPUTS:

Amp/Booth:...0 dB 1 V, 400 Ohm Max:...20 V Peak-to-Peak Rec...150 mV, 5 KOhm Zone...0 dB 1 V 400 Ohm Balanced...6 dB 2 V 400 Ohm

GENERAL:

Frequency Response:..20 Hz - 20 KHz +/- 2 dB Distortion:...0.02% S/N Ratio:..Better Than 80 dB Talkover Attenuation:...-16 dB Headphone Impedance:...16 Ohm Power Source:..115/230 V, 60/50 Hz, 20 W Unit Dimensions:...W 19" x H 4" x D 9.1"...(482.6 x 100.6 x 231.2 mm) Weight:.... 10.34 lbs (4.7 kg)

REMARQUE: GEMINI, DANS LE CADRE D'UN SOUCI CONSTANT D'AMELIORATION DE SES PRODUITS, SE RESERVE LE DROIT DE LES MODIFIER SANS AUCUN PREAVIS.

ENTREES:

SORTIES:

NOTES:

IN THE USA: IF YOU EXPERIENCE PROBLEMS WITH THIS UNIT, CALL 1-732-738-9003 FOR GEMINI CUSTOMER SERVICE. DO NOT ATTEMPT TO RETURN THIS EQUIPMENT TO YOUR DEALER.
Parts of the design of this product may be protected by worldwide patents. Information in this manual is subject to change without notice and does not represent a commitment on the part of the vendor. Gemini Sound Products Corp. shall not be liable for any loss or damage whatsoever arising from the use of information or any error contained in this manual. No part of this manual may be reproduced, stored in a retrieval system or transmitted, in any form or by any means, electronic, electrical, mechanical, optical, chemical, including photocopying and recording, for any purpose without the express written permission of Gemini Sound Products Corp. It is recommended that all maintenance and service on this product is performed by Gemini Sound Products Corp. or its authorized agents. Gemini Sound Products Corp. will not accept liability for loss or damage caused by maintenance or repair performed by unauthorized personnel.
Worldwide Headquarters 120 Clover Place, Edison, NJ 08837 USA Tel: (732) 738-9003 Fax: (732) 738-9006
France Gemini France (GSL) 1, Allee dEffiat, Parc de levnement, 91160 Longjumeau, France Tl: + Fax: + Germany Gemini Sound Products GmbH Liebigstr. 16, Haus B - 3.0G, 85757 Karlsfeld, Germany Tel: 08131 - 39171-0 Fax: 08131 - 39171-9 UK Gemini Sound Products Unit C4 Hazleton Industrial Estate, P08 9JU Waterlooville, UK Tel: 00880 Fax: 00990 Spain Gemini Sound Products S.A. Rosello, 516, 08026 Barcelona, Spain, Tel: 349-3435-0814 Fax: 3493-347-6961
___________________________________________________ Gemini Sound Products Corp. 2004 All Rights Reserved.