H. Balakrishnan, S. Seshan, M. Stemm, and R. Katz. Analyzing Stability in Wide-Area Network Performance. In Proc. ACM SIGMETRICS, June 1997.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....are captured well by the Lognormal distribution with = 0.005, # = 1.55 (transfer time) and = 9.35, # = 1.6 (throughput) respectively. Our finding of throughput coincides with earlier observations made by Balakrishnan et al. using traces from the 1996 Atlanta Summer Olympic Games web server [7]. The bursts at 3:00am on 02 16, 12:00pm on 02 25, and 1:00pm on 02 25 occur because of backup operations and a restart of the session manager respectively. 12 To identify the primary contributor to network transfer time, we again computed the correlation coefficient between transfer times ....

H. Balakrishnan, S. Seshan, M. Stemm, and R. H. Katz. Analyzing stability in wide-area network performance. Proc. of ACM SIGMETRICS Conference on Measurement and Modeling of Computer Systems, June 1997.

....clients experience widely different throughput, from 1 Kbps up to 10 Mbps. It is evident that the Web server was not the bottleneck of the endto end performance for most clients. Meanwhile it re confirms the fact that end to end performance varies significantly for different hosts, as reported in [3]. In addition, 33.8 of the clients have throughput less than 20 Kbps, i.e. less than stereo quality audio or video encoding rate. This suggests that the Internet path can often become a bottleneck for real time applications. For example, if an IP based wireless phone call is made from a mall to ....

....variation within a factor of 4 during the trace period; when considering the variation between the maximum and mean throughput, over 90 of the hosts have throughput variation within a factor of 2. These results are consistent with the previous studies, which reported similar degree of stability [3, 21]. On the other hand, throughput for some hosts vary by up to three orders of magnitude. Next, we examine how long a client s throughput remains stable. We use the notion of operational throughput constancy introduced in [21] We say that the throughput remains operationally constant in a region ....

[Article contains additional citation context not shown here]

H. Balakrishnan, S. Seshan, M. Stemm, and R. H. Katz. Analyzing Stability in Wide-Area Network Performance. In Proceedings of ACM SIGMETRICS '97, 1997.

....are captured well by the Lognormal distribution with tt 0.005, cr 1.55 (transfer time) and tt 9.35, cr 1.6 (throughput) respectively. Our finding of throughput coincides with earlier observations made by Balakrishnan et al. using traces from the 1996 Atlanta Summer Olympic Games web server [7]. To identify the primary contributor to network transfer time, we again computed the correlation coefficient between transfer times and document sizes. The result, 0.0031, reveals that in general there is no clear correlation between the two. A stronger correlation was observed when we ....

H. Balakrishnan, S. Seshan, M. Stemm, and R. H. Katz. Analyzing stability in wide-area network performance. Proc. of ACM SIGMETRICS Conference on Measurement and Modeling of Computer Systems, June 1997.

....in Fig. 8, each component may be modeled as a multi user system in Fig. 6. Note that a user shares the system resources with different users in different components. Depending on the location of the user and the server, the Internet may look different in terms of delay, capacity, and load [8]. In general, it is relatively easy to obtain these network characteristics for the Intranet and the connection between the Intranet and the Internet backbone. Systems like the one proposed in [9] may help us getting information about backbone network and Web servers b s T B T B B i T i k ....

....and may vary with time. However, in general, it does indicate the quality of the link at that time. If the user has downloaded some pages from the server recently, say in the last 10 minutes, the average speed at which the files were downloaded is a better estimation of the actual link condition [8]. This link quality information is provided to the user through an icon placed next to the corresponding link. This can potentially reduce traffic on congested links, because most people would choose a faster link if several options are available. Figure 10 shows the icons used in our program. ....

H. Balakrishnan al., "Analyzing Stability in Wide-Area Network Performance, " Proc. of ACM SIGMETRICS Conf. Measurement & Modeling of Comp. Sys., Seattle, WA, June 1997.

....to clients. This work also builds on earlier research work that has examined Web performance from the viewpoint of individual improvements in reducing user perceived latency or load on the network. The set of ideas includes compression and delta encoding [14] stability in network performance [4], examining impact of various protocol variations of HTTP [16, 10] and bundling resources [22] What these works have in common is the use of a single idea to explore impact on Web performance. Each of these pieces of research differ in their evaluation environment in the sense that they use ....

H. Balakrishnan, M. Stemm, S. Seshan, and R. H. Katz. Analyzing Stability in Wide-Area Network Performance. In Measurement and Modeling of Computer Systems, pages 2--12, 1997.

....episodes, and not from nearby losses. Throughput has a close coupling with the loss process, and can often be modeled as a stationary IID process for a period of hours. Several studies have also examined similar issues by studying traces gathered passively using a packet sniffer. The authors in [2] used traces from the 1996 Olympic Games Web site to analyze the spatial and temporal stability of TCP throughput. Using traceroute data, they constructed a tree rooted at the server and extending out to the client hosts. Clients were clustered based on how far apart they were in the tree. The ....

....In [1] Allman uses traces gathered at the NASA Glenn Research Center Web server to study issues such as TCP and HTTP option usage, RTT and packet size distributions, etc. Mogul et al. 12] uses packet level traces to study the effectiveness of delta compression for HTTP. Our study is similar to [2] in that it is based on packet traces gathered passively at a busy server. However, our analysis is different in many ways. We focus on packet loss rate rather than TCP throughput for the reasons mentioned previously. More importantly, we go beyond simply characterizing the end to end loss rate ....

H. Balakrishnan, S. Seshan, M. Stemm, and R. H. Katz. Analyzing Stability in Wide-Area Network Performance. In Proceedings of ACM SIGMETRICS'98, June 1997.

....in larger time scales and the presence of a de nite diurnal cycle. Arlitt and Williamson [2] observed similar regularity in client activity levels in web servers. Balakrishnan et al. observed regularity in their in depth analysis of trac to the web server of the Atlanta Summer Olympic Games [3]. These studies suggest predictable patterns in the e ectiveness level of the cache. Most cache replacement policies evict a page which maximizes (or minimizes) a particular eviction function. LRU (least recently used) for example, evicts the page with largest passed time since its last request, ....

H. Balakrishnan, S. Seshan, Stemm M., and Katz R. H. Analyzing stability in wide-area network performance. In Proceedings of the ACM SIGMETRICS '97 Conference, 1997. http://HTTP.CS.Berkeley.edu/~stemm/pubs.html.

....episodes, and not from nearby losses. Throughput has a close coupling with the loss process, and can often be modeled as a stationary IID process for periods of hours. Several studies have also examined similar issues by studying traces gathered passively using a packet sniffer. The authors in [2] used traces from the 1996 Olympic Games Web site to analyze the spatial and temporal stability of TCP throughput. Using traceroute data, they constructed a tree rooted at the server and extending out to the client hosts. Clients were clustered together based on how far apart they are in the tree. ....

....In [1] Allman uses traces gathered at the NASA Glenn Research Center Web server to study issues such as TCP and HTTP option usage, RTT and packet size distributions, etc. Mogul et al. uses packet level traces to study the effectiveness of delta compression for HTTP [15] Our study is similar to [2] in that it is based on packet sniffer traces gathered passively at a busy server. However, our analysis is different in many ways. We focus on packet loss rate rather than TCP throughput for the reasons mentioned previously. Our analysis of spatial locality considers operationally meaningful ....

H. Balakrishnan, S. Seshan, M. Stemm, and R. H. Katz. Analyzing stability in wide-area network performance. In Proceedings of ACM SIGMETRICS '98, June 1997.

....are captured well by the Lognormal distribution with = 0.005, # = 1.55 (transfer time) and = 9.35, # = 1.6 (throughput) respectively. Our finding of throughput coincides with earlier observations made by Balakrishnan et al. using traces from the 1996 Atlanta Summer Olympic Games web server [7]. To identify the primary contributor to network transfer time, we again computed the correlation coefficient between transfer times and document sizes. The result, 0.0031, reveals that in general there is no clear correlation between the two. A stronger correlation was observed when we ....

H. Balakrishnan, S. Seshan, M. Stemm, and R. H. Katz. Analyzing stability in wide-area network performance. Proc. of ACM SIGMETRICS Conference on Measurement and Modeling of Computer Systems, June 1997.

....the TCP sender detects a loss during congestion avoidance. Measurement study of Internet traces shows that the WAN performance is reasonably stable over terms of several minutes; meanwhile, nearby hosts experience similar or identical throughput performance within a time period measured in minutes [30, 9, 37]. Such level of stability suggests sharing performance information both temporarily and spatially (across many co located hosts) can help to more accurately determine network performance, in particular, a connection s fair share of network resources (Wc ) 5. TCP SPAND: SYSTEM DESIGN AND ....

....sharing; on the other hand, a large S means the performance gateway needs to keep a large amount of state; in addition, the choice of S needs to match the level of stability reported in the literature. Currently, we set S to 5 minutes. It clearly matches the level of stability reported in [30, 9, 37], which is a few minutes. Below we use MSNBC traces to demonstrate that a 5 minute sliding window can achieve significant sharing while only requiring moderate amount of state kept by the performance gateway. Figure 6 shows the cumulative distribution of the time between two consecutive requests ....

H. Balakrishnan, S. Seshan, M. Stemm, and R. Katz, "Analyzing Stability in Wide-Area Network Performance," Proc. SIGMETRICS '97, 1997.

....tools to merge distributed data into a coherent model. Measurement study of Internet traces shows that the WAN performance is reasonably stable over terms of several minutes; meanwhile, nearby hosts experience similar or identical throughput performance within a time period measured in minutes [5, 45]. Our model parameterization tool outputs simulation model at the time scale of tens of minutes for hour long traffic, which matches the level of stability reported in previous study and hence is applicable to simulate present traffic and predict short term traffic trend. However, to simulate and ....

Hari Balakrishnan, Mark Stemm, Srinivasan Seshan, and Randy H. Katz. Analyzing stability in wide-area network performance. In SIGMETRICS/Performance, 1997.

....terms of user perceived response time. C. Related Work Several researchers have addressed performance issues related to TCP and the Web. Padmanabhan and Katz [23] proposed a fast start mechanism for TCP that increases the initial TCP window size (based on cached connection state information [6]) combined with a low packet priority for these extra packets at TCP connection startup. The goal of this approach is to decrease the number of round trip times incurred, particularly for short lived Web document transfers. The fast start mechanism requires Internet routers to consider packet ....

H. Balakrishnan, S. Seshan, M. Stemm, and R. Katz, "Analyzing Stability in Wide-Area Network Performance", Proceedings of ACM SIGMETRICS, Seattle, WA, pp. 2-12, June 1997.

....of the avail bw. Paxson s metric # is fairly predictable: on average, a measurement of # at a given path falls within 10 of later # measurements for periods that last for several hours [36] Balakrishnan et.al. examined the throughput stationarity of successive Web transfers to a set of clients [5]. The throughput to a given client appeared to be piecewise stationary in timescales that extend for hundreds of minutes. Additionally, the throughput of successive transfers to a given client varied by less than a factor of 2 over three hours. A more elaborate investigation of the stationarity of ....

H. Balakrishnan, S. Seshan, M. Stemm, and R. H. Katz. Analyzing Stability in Wide-Area Network Performance. In Proceedings of ACM SIGMETRICS, June 1997.

....flow size distribution [14] This result was consistent with observation in [10] that a small number of flows accounted for a significant number of the total bytes. In recent work, Sarvotham et al. [20] found that a single high rate flow usually accounts for the burstiness in aggregate tra#c. In [2], the authors look at the distribution of throughput across connections between hosts and a web server and find that the rates are often consistent with a log normal distribution. These papers have all made important observations. In this paper, we aim to go beyond this previous work, looking at ....

....packet traces. One possible explanation of this di#erence is that file sizes are potentially unbounded while flow rates are constrained by link bandwidths. A previous study of rate distributions at a web server suggested that the rate distributions were well described by a log normal distribution [2]. To test that hypothesis, we use the quantile quantile plot (Q Q plot) 3] to compare the flow rate distribution with analytical models. The Q Q plot determines whether a data set has a particular theoretical distribution by plotting the quantiles of the data set against the quantiles of the ....

[Article contains additional citation context not shown here]

H. Balakrishnan, S. Seshan, M. Stemm, and R. Katz, "Analyzing Stability in Wide-Area Network Performance," In Proc. ACM SIGMETRICS' 97, June 1997.

....to our present study and discusses the benefits of working with shared, passive measurements from a collection of hosts rather than with active measurements from a single host when trying to determine network path characteristics. In this context, another relevant study is by Balakrishnan et al. [6], whose empirical investigations revealed significant stability of network paths characteristics (i.e. available bandwidth) over time and across hosts that are close to each other; in other words, hosts which share portions of a network path tend to obtain similar amounts of throughput. Stability ....

H. Balakrishnan, S. Seshan, M. Stemm, and R. H. Katz. Analyzing stability in wide-area network performance. Proc. of the ACM/SIGMETRICS'97, Seattle, WA, 1997.

.... Another approach is to use prefetching to reduce response times, by hiding server and network latency [6, 24] A complementary approach is to make the Web server more powerful through the use of a clustered architecture, in which multiple machines function cohesively as a single Web server (e.g. [1, 3, 9, 12, 13, 14, 15, 18, 19, 25, 30]) Such an architecture offers multiple benefits: first, more machines means more capacity to handle requests; second, reliability and availability are improved since the server can continue to operate should some machines go down; third, a consistent external interface can be preserved; and ....

H. Balakrishnan, S. Seshan, M. Stemm, and R. Katz, "Analyzing Stability in Wide-Area Network Performance", Proceedings of the 1997 ACM SIGMETRICS Conference, Seattle, WA, pp. 2-12, June 1997.

....in larger time scales and the presence of a definite diurnal cycle. Arlitt and Williamson [2] observed similar regularity in client activity levels in web servers. Balakrishnan et al. observed regularity in their in depth analysis of traffic to the web server of the Atlanta Summer Olympic Games [3]. These studies suggest predictable patterns in the effectiveness level of the cache. Most cache replacement policies evict a page which maximizes a particular eviction function. lru (least recently used) for example, evicts the page with largest passed time since its last request, lfu (least ....

H. Balakrishnan, S. Seshan, Stemm M., and Katz R. H. Analyzing stability in wide-area network performance. In Proceedings of the ACM SIGMETRICS '97 Conference, 1997. http://HTTP.CS.Berkeley.edu/~stemm/pubs.html.

....models assume arbitrarilylong TCP connections, where the complex, random behavior of TCP can be summarized accurately by deriving the expected bandwidth in the steady state. However, a wealth of evidence suggests that most TCP connections today are short usually transferring under 10KB [CBC95, BSSK97, Mah97a, GB97, TMW97] Those models that consider shorter TCP connections assume that there are no lost packets. This is an unrealistic assumption in an Internet where loss rates are often around 5 [Pax97, BPS 98] Hence it is important to understand how well the steady state, lossy models ....

.... new TCP connection for each object embedded in a web page [BLFF96, Ste96] Several studies have found that this results in TCP connections with a median size of 2 3KB, an average transfer size of 8 12KB, with the vast majority of connections transferring less than 10KB, BC94, CBC95, Mog95, Ste96, BSSK97, Mah97a, GB97, TMW97] The vast majority of objects are smaller than the average because of the heavy tailed distribution of object sizes. These flows are quite short in TCP terms with a typical MSS of 1460 bytes, 10KB transfers are 7 segments long. HTTP 1.1 and some implementations of HTTP ....

[Article contains additional citation context not shown here]

Hari Balakrishnan, Mark Stemm, Srinivasan Seshan, and Randy H. Katz. Analyzing stability in wide-area network performance. In Proceedings of the 1997 SIGMETRICS Conference, June15--18 1997.

....a new technique, which we call TCP SPAND, to improve TCP start up performance. We first theoretically analyze the problem and derive optimal TCP initial parameters for the simplest scenario in which there is only a single TCP connection and the network condition is given. Recent studies [18] [7] have shown nearby hosts experience similar or identical network performance within a time period measured in minutes. Such level of network stability allows us to abstract an end to end communication path as a virtual link, which is quite stable, and apply our optimality results to it. In order ....

....of a virtual link by observing its performance. However measurements from a single host can sometimes be redundant, inaccurate, or out of date. The study of Internet traces shows nearby hosts experience similar or identical throughput performance within a time period measured in minutes [18] [7]. This suggests sharing performance information across hosts can help to determine the network performance more accurately. Based on the observation, 22] proposed an architecture called SPAND (Shared PAssive Network Discovery) that determines network characteristics by making shared, passive ....

H. Balakrishnan, S. Seshan, M. Stemm, and R. H. Katz. Analyzing Stability in Wide-Area Network Performance. In Proc. ACM SIGMETRICS '97, 1997.

....the TCP sender detects a loss during congestion avoidance. Measurement study of Internet traces shows that the WAN performance is reasonably stable over terms of several minutes; meanwhile, nearby hosts experience similar or identical throughput performance within a time period measured in minutes [25, 7, 31]. Such level of stability suggests sharing performance information both temporarily and spatially (across many co located hosts) can help to more accurately determine network performance, in particular, a connection s fair share of network resources (Wc ) 5. TCP SPAND: SYSTEM DESIGN AND ....

....sharing; on the other hand, a large S means the performance gateway needs to keep a large amount of state; in addition, the choice of S needs to match the level of stability reported in the literature. Currently, we set S to 5 minutes. It clearly matches the level of stability reported in [25, 7, 31], which is a few minutes. Below we use MSNBC traces to demonstrate that a 5 minute sliding window can achieve significant sharing while only requiring moderate amount of state kept by the performance gateway. Figure 6 shows the cumulative distribution of the time between two consecutive requests ....

H. Balakrishnan, S. Seshan, M. Stemm, and R. Katz, "Analyzing Stability in Wide-Area Network Performance," Proc. SIGMETRICS '97, 1997.

....a round robin manner. In a random dispatching scheme, the packets are forwarded to one of the randomly chosen routes. In the hash based scheme, the route is chosen by hashing on the source address and the port number. We use the tcpdump traces collected from the Official 1996 Olympic Web Server [2] as the traffic profile for this experiment. Traces were collected for approximately 2 weeks in 20minute segments. The entire trace consists of approximately 60 million HTTP requests, from about 725,000 clients. In figure 3 we plot the distribution of connection holding time and amount of data ....

H. Balakrishnan, M. Stemm, S. Seshan, and R. H. Katz. Analyzing stability in wide-area network performance. In SIGMETRICS'97, Seattle, June 1997.

.... Another approach is to use prefetching to reduce response times, by hiding server and network latency [6, 18] A complementary approach is to make the Web server more powerful through the use of a clustered architecture, in which multiple machines function cohesively as a single Web server (e.g. [1, 3, 8, 9, 10, 11, 12, 14, 15, 19, 24]) Such an architecture offers multiple benefits: first, more machines means more capacity to handle requests; second, reliability and availability are improved since the server can continue to operate should some machines go down; third, a consistent external interface can be preserved; and ....

H. Balakrishnan, S. Seshan, M. Stemm, and R. Katz, "Analyzing Stability in Wide-Area Network Performance", Proceedings of the 1997 ACM SIGMETRICS Conference, Seattle, WA, pp. 2-12, June 1997. 168 Achieving Load Balance and Effective Caching in Clustered Web Servers

....maximum bandwidth, available server bandwidth, nor dynamic estimates of available bandwidth between client and server that may be obtainable from the monitoring of on demand HTTP traffic. Available bandwidth can be estimated from TCP [19] with good accuracy for time periods of a few minutes [1] to a few hours [17] Thus, a client will not necessarily 2 A low priority datagram is basically filler traffic for the purpose of this paper. It can be implemented as an IPv6 datagram with priority 0 or 1, or, to some extent, as an IPTOS LOWCOST IPv4 datagram. Trace Date Requests Hosts Objects ....

Hari Balakrishnan, Srinivasan Seshan, Mark Stemm, and Randy H. Katz. Analyzing stability in wide-area network performance. In Proceedings of ACM SIGMETRICS Conference on Measurement and Modeling of Computer Systems (SIGMETRICS 97), Seattle WA, USA, June 1997.

....use of multiple network devices for both availability and e#ciency reasons. Multiple interfaces were not available at any point in time, just the best interface that is selected according to a specific policy. Goals similar to those of Stanford s MosquitoNet, InfoPad and Daidalus project (e.g. [12]) were also discussed in [13] While these groups focused more on mobile IP implementations, J. Inouye et al. [13] were dealing more with dynamic reconfiguration policies. There is a large amount of work focused on routing protocols to support mobility and some on ad hoc mobile networks [14] ....

H. Balakrishnan, S. Seshan, M. Stemm, and R. H. Katz, "Analyzing stability in wide-area network performance," SIGMETRICS Conference on Measurement and Modeling of Computer Systems, June 1997.

....reasonable for future Integrated Services Networks. For instance, the bandwidth requirement of MPEG video stream is between 1 6 Mbps, which is reasonably small as compared to link speeds of 155 Mbps to 2:4 Gbps. The average throughput of a typical web connection is also small (4 Kbps to 128 Kbps) [22] as compared to the link speeds. Though the holding time for web connections is shown to be heavy tailed [23] most of the connections are still short lived. The session arrival process is also not as bursty as the data arrival process. We therefore simulated using the most simplistic model. We ....

Hari Balakrishnan, Srinivasan Seshan, Mark Stemm, and Randy H. Katz, "Analyzing stability in wide-area network performance," in Proceedings of ACM SIGMETRICS Conference on Measurement and Modeling of Computer Systems, June 1997.

....bulk transfer throughput. While these models are very successful at predicting steady state throughput [5, 38] many recent studies have noted that the majority of TCP flows traveling over the wide area Internet are very short, with mean sizes around 10KB and median sizes less than 10KB [11, 4, 23, 16, 41, 10]. Because these flows are so short, they often spend their entire lifetime in TCP s slow start mode, without suffering a single loss. Since the steady state models assume flows suffer at least one loss, they are undefined for this common case. The second class of models focuses on these short ....

Hari Balakrishnan, Mark Stemm, Srinivasan Seshan, and Randy H. Katz. Analyzing stability in wide-area network performance. In SIGMETRICS '97, June 1997.

No context found.

H. Balakrishnan, S. Seshan, M. Stemm, and R.H. Katz. Analyzing Stability in Wide-Area Network Performance. In Proc. ACM SIGMETRICS, pages 2--12, Seattle, WA, June 1997.

No context found.

H. Balakrishnan, S. Seshan, M. Stemm, and R. H. Katz. Analyzing stability in wide-area network performance. In Proceedings of ACM SIGMETRICS, Seattle, WA, June 1997.

No context found.

H. Balakrishnan, S. Seshan, M. Stemm, and R. H. Katz. Analyzing stability in wide-area network performance. In Proceedings of ACM SIGMETRICS, Seattle, WA, June 1997.

No context found.

H. Balakrishnan, S. Seshan, M. Stemm, and R.H. Katz. Analyzing Stability in Wide-Area Network Performance. In Proc. ACM SIGMETRICS '97, June 1997.

No context found.

BALAKRISHNAN, H., SESHAN, S., STEMM, M., AND KATZ, R. Analyzing Stability in Wide-Area Network Performance. In Proc. ACM SIGMETRICS (Seattle, WA, June 1997), pp. 2--12.

....individually: ############ #### ###### ## # ######## # #. RON does not attempt to find optimal throughput paths, but strives to avoid paths of low throughput when good alternatives are available. Given the time varying and somewhat unpredictable nature of available bandwidth on Internet paths [2, 19], we believe this is an appropriate goal. From the standpoint of improving the reliability of path selection in the face of performance failures, avoiding bad paths is more important than optimizing to eliminate small throughput differences between paths. While a characterization of the utility ....

....available bandwidths may help determine a good path selection threshold, we believe that more than a 50 bandwidth reduction is likely to reduce the utility of many programs. This threshold also falls outside the typical variation observed on a given path over time scales of tens of min utes [2]. We therefore concentrate on avoiding throughput faults of this order of magnitude. Throughput intensive applications typically use TCP or TCPlike congestion control, so the throughput optimizer focuses on this type of traffic. The performance of a bulk TCP transfer is a function of the ....

BALAKRISHNAN, H., SESHAN, S., STEMM, M., AND KATZ,R. Analyzing Stability in Wide-Area Network Performance. In Proc. ACM SIGMETRICS (Seattle, WA, June 1997), pp. 2--12.

....individually: lossratepath = 1 l2path (1 lossrate l ) RON does not attempt to find optimal throughput paths, but strives to avoid paths of low throughput when good alternatives are available. Given the time varying and somewhat unpredictable nature of available bandwidth on Internet paths [2, 19], we believe this is an appropriate goal. From the standpoint of improving the reliability of path selection in the face of performance failures, avoiding bad paths is more important than optimizing to eliminate small throughput differences between paths. While a characterization of the utility ....

....available bandwidths may help determine a good path selection threshold, we believe that more than a 50 bandwidth reduction is likely to reduce the utility of many programs. This threshold also falls outside the typical variation observed on a given path over time scales of tens of min utes [2]. We therefore concentrate on avoiding throughput faults of this order of magnitude. Throughput intensive applications typically use TCP or TCPlike congestion control, so the throughput optimizer focuses on this type of traffic. The performance of a bulk TCP transfer is a function of the ....

BALAKRISHNAN, H., SESHAN, S., STEMM, M., AND KATZ, R. Analyzing Stability in Wide-Area Network Performance. In Proc. ACM SIGMETRICS (Seattle, WA, June 1997), pp. 2--12.

....individually: lossrate path = 1 l2path (1 lossrate l ) RON does not attempt to find optimal throughput paths, but strives to avoid paths of low throughput when good alternatives are available. Given the time varying and somewhat unpredictable nature of available bandwidth on Internet paths [1, 37], we believe this is an appropriate goal. From the standpoint of improving the reliability of path selection in the face of performance failures, avoiding bad paths is more important than optimizing to eliminate small bandwidth differences between paths. While a characterization of the utility ....

.... the utility received by programs at different bandwidth rates, similar to that developed by Breslau and Shenker [3] is beyond the scope of this work, we believe that more than a 50 difference is likely to reduce the utility of many programs, and also falls outside the typical variation observed [1]. We concentrate on avoiding throughput faults on this order of magnitude. The performance of bulk TCP transfers is a function of the connection s round trip latency and 47 the packet loss rate it observes. Throughput optimization combines the latency and loss metrics using a simplified version ....

H. Balakrishnan, S. Seshan, M. Stemm, and R.H. Katz. Analyzing Stability in Wide-Area Network Performance. In Proc. ACM SIGMETRICS '97, pages 2--12, Seattle, WA, June 1997.

....into the network. We show later that this unnecessary traffic can quickly grow to become a non negligible part of the traffic reaching a busy Web server. Probing from a single host prevents a client from using the past information of nearby clients to predict future performance. Recent studies [2][15] have shown that performance from a client to a server is stable for many minutes and is identical to the performance observed by other nearby clients. In this paper, we show examples where using shared rather than isolated information increases the likelihood that previously collected network ....

....that is not useful to any application. For example, pathchar sends (at least) tens of kilobytes of probe traffic per hop, and cprobes send 6 kilobytes per cprobe. This amount of probe traffic is a significant fraction (approximately 20 ) of the mean transfer size for many WWW connections ( 1] [2]) as well as a significant fraction of the mean transfer size for many WWW sessions. 2.2 Server Selection Systems The primary application of the of network probing algorithms described in the previous section has been in dynamic wide area server selection. 4] uses cprobes and bprobes to select ....

[Article contains additional citation context not shown here]

H. Balakrishnan, S. Seshan, M. Stemm, and R.H. Katz. Analyzing Stability in Wide-Area Network Performance. In Proc. ACM SIGMETRICS '97, June 1997.

....can quickly grow to become a non negligible part of the traffic reaching busy Web servers, reducing their efficiency and sometimes their scalability. Probing from a single host prevents a client from using the past information of nearby clients to predict future performance. Recent studies [3] [34] have shown that network performance from a client to a server is often stable for many minutes and very similar to the performance observed by other nearby clients, so there are potential benefits of sharing information between hosts. In Section 3.2, we show examples where using shared ....

....is not useful to any application. For example, pathchar sends at least tens of kilobytes of probe traffic per hop, and a cprobe sends 6 kilobytes of traffic per probe. This amount of probe traffic is a significant fraction (approximately 20 ) of the mean transfer size for many Web connections ( 2] [3] [16] as well as a significant fraction of the mean transfer size for many Web sessions. In contrast, SPAND introduces a minimal amount of additional traffic in the form of small performance reports and queries. We discuss in more detail the limitations of active probing in Section 3.3. How ....

[Article contains additional citation context not shown here]

H. Balakrishnan, S. Seshan, M. Stemm, and R.H. Katz. Analyzing Stability in Wide-Area Network Performance. In Proc. ACM SIGMETRICS '97, June 1997.

....receiver buffer by default, and applications such as Web browsers often do not explicitly request a larger buffer. On the other hand, the congestion window tends to be small because the individual connections tend to be short in length; the mean connection length is 5.5 KB and the median 3. 1 KB [7]. As a result, slow start fails to build up a large window. The tendency connections to be short in length also implies that slow start dominates the congestion avoidance phase. As indicated in Table 4.1, an overwhelming majority (85 ) of the packets were transmitted by the server during the slow ....

....8.4) show that fast start is quite successful in meeting these goals. While fast start is equally applicable to a standard TCP connection as well as to a TCP session, for ease of exposition we present our discussion in terms of a single TCP connection. 8.2. 1 Basic Idea Various studies (e.g. [7], 88] have shown that the available network bandwidth tends to be stable for periods ranging from a few minutes to tens of minutes. Since the length of a typical pause during an active Web browsing session tends to be much shorter (for instance, 65] reports a median of 15 seconds) it may be ....

[Article contains additional citation context not shown here]

H. Balakrishnan, S. Seshan, M. Stemm, and R.H. Katz. Analyzing Stability in Wide-Area Network Performance. In Proc. ACM SIGMETRICS '97, June 1997. 209

No context found.

H. Balakrishnan, S. Seshan, M. Stemm, and R. Katz. Analyzing Stability in Wide-Area Network Performance. In Proc. ACM SIGMETRICS, June 1997.

No context found.

Hari Balakrishnan, Mark Stemm, Srinivasan Seshan, and Randy H. Katz. Analyzing stability in wide-area network performance. In SIGMETRICS '97, June 1997.

No context found.

H. Balakrishnan and S. Srinivasan 1997. Analyzing stability in wide-area network performance. ACM SIGMETRICS Conference on Measurement and Modeling of Computer Systems, 1997.

No context found.

H. Balakrishnan, M. Stemm, S. Seshan, and R. H. Katz. Analyzing Stability in Wide-Area Network Performance. In Proc. of the ACM SIGMETRICS, Seattle, WA, pages 2--12, 1997.

No context found.

BALAKRISHNAN, H., SESHAN, S., STEMM, M., AND KATZ, R. H. Analyzing Stability in Wide-Area Network Performance. In ACM SIGMETRICS (June 1997). 11

No context found.

H. Balakrishnan, M. Stemm, S. Seshan, and R. H. Katz, "Analyzing stability in wide-area network performance," in SIGMETRICS/Performance, 1997.

No context found.

BALAKRISHNAN, H., SESHAN, S., STEMM, M., AND KATZ, R. H. Analyzing Stability in Wide-Area Network Performance. In ACM SIGMETRICS (June 1997).

No context found.

H. Balakrishnan, M. Stemm, S. Seshan, and R.H. Katz, "Analyzing stability in wide-area network performance," Proceedings of the ACM SIGMETRICS'97, pp. 2--12, 1997

No context found.

Balakrishnan Hari and Srinivasan Seshan. Analyzing stability in wide-area network performance. ACM SIGMETRICS Conference on Measurement and Modeling of Computer Systems, 1997.

No context found.

H. Balakrishnan, M. Stemm, S. Seshan, and R. H. Katz. Analyzing Stability in Wide-Area Network Performance. In Proc. of the ACM SIGMETRICS, Seattle, WA, pages 2--12, 1997.

No context found.

H. Balakrishnan, S. Seshan, M. Stemm, and R. H. Katz. Analyzing stability in wide-area network performance. In ACM SIGMETRICS Conference Proceedings, Seattle WA, USA, June 1997. ACM.

No context found.

Hari Balakrishnan, Mark Stemm, Srinivasan Seshan, and Randy H. Katz. Analyzing stability in wide-area network performance. In SIGMETRICS/Performance, 1997.

No context found.

Balakrishnan Hari and Srinivasan Seshan. Analyzing stability in wide-area network performance. ACM SIGMETRICS Conference on Measurement and Modeling of Computer Systems, 1997.

First 50 documents  Next 50

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC