IMR-Pathload: Robust Available Bandwidth Estimation Under End-Host Interrupt Delay

Abstract

Abstract. Many paths in PlanetLab cannot be measured by Pathload. One of the main reasons for this is timing irregularities caused by interrupt moderation of network hardware, which delays generation of interrupts for a certain period of time to reduce per-packet CPU overhead. Motivated by this problem, we study Pathload in detail under various end-host interrupt delays and find that its trend detection mechanism becomes susceptible to non-negligible interrupt delays, making it unable to measure network paths under such conditions. To overcome this, we propose a new method called IMR-Pathload (Interrupt Moderation Resilient Pathload ), which incorporates robust trend detection algorithms based on signal de-noising techniques and reliably estimates available bandwidth of network paths under a wide range of interrupt delays. Through experiments in Emulab and Internet, we find that IMR-Pathload substantially improves Pathload's measurement reliability and produces accurate bandwidth estimates under a variety of real-life conditions. Key words: Bandwidth estimation, network measurement, interrupt moderation, and interrupt delays IntroductionBandwidth of Internet paths is an important metric for applications. Extensive research has been conducted over the years and the vast majority of work in this area focuses on end-to-end measurement. Although several techniques [4], [13], [11], [12], [14] attempt to measure capacity of the narrow link (i.e., the slowest link in a path) or both capacity and available bandwidth of the tight link (i.e., link with the smallest available bandwidth over a path), many measurement techniques and public tools (such as [6], [9], [16]) have been developed to estimate available bandwidth of the tight link. These methods mainly focus on fast estimation with high accuracy under a various traffic conditions. However, since the ultimate goal of bandwidth estimators is to measure diverse Internet paths, before being a full-blown measurement tool, it is highly desirable that tools are resilient to timing irregularities caused by various OS scheduling delay jitter or hardware interrupt moderation in real networks.Note that to accurately measure bandwidth, all existing methods heavily rely on high-precision delay measurement of probe packets at end-hosts. However, irregular timing due to interrupt moderation at network interface cards (NICs) has been identified as the major problem of existing bandwidth estimation tools in practice [15]. To reduce the effect of interrupt moderation, recent tools such as Pathchirp [16] and Pathload described in [15] incorporate mechanisms that aim to "weed out" packets affected by interrupt delays. However, Pathchirp requires manual modification to force it to send (often substantially) more probing packets to obtain an accurate estimate, prolonging measurement undesirably. On the other hand, Pathload attempts to filter out affected packets without increasing the number of probing packets, which unfortunately has a limited effect when interrupt delays become non-t...

Characterizing tight-link bandwidth of multi-hop paths using probing response curves

“…This makes Pathload's estimation much more susceptible to error, which happens fairly often in practice. Kang et al [15] overcome this problem in a recent proposal called Interrupt Moderation Resilient Pathload (IMR-Pathload), which utilizes signal de-noising techniques such as wavelet decomposition and window-based averaging in detecting a delay trend exists in one-way delay samples of probe packets.…”

“…We assess the performance of PRC-MT in scenarios with non-negligible interrupt delays in comparison with Pathload [26], IMR-Pathload [15], Pathchirp [27], IGI/PTR [6] using metric A t and Pathrate [3], CapProbe [16] using metric C t when the narrow link coincides with the tight link. For available bandwidth A t , our results show that PRC-MT exhibits no negative side-effects related to interrupt moderation, converges in 90 − 140 seconds in all examined topologies, and outperforms the other studied methods in terms of accuracy (1−5% error).…”

“…Among tools evaluated with no interrupt delay (δ = 0) in Section III, use of interrupt moderation affects Pathload the most, while others exhibit estimation accuracy that is similar to that in the cases without using interrupt moderation. Recently, Kang et al [15] proposed a measurement tool called IMR-Pathload that is resilient to various interrupt delays and significantly improves Pathload's estimation reliability under such conditions (see [15] for detailed behavior of Pathload and IMR-Pathload under a wide range of interrupt delays). We include this method in comparison of estimation tools under interrupt moderation in the following subsections.…”

“…Notice that IMR-Pathload's estimation accuracy is as good as PRC-MT, but Pathload is unable to produce estimates for any of the cases as shown in the table as empty cells. This suggests that Pathload's algorithm is susceptible to non-trivial interrupt delays (see [15] for details).…”

“…One direction of this prior work is to measure capacity of the tight link, which utilizes inter-packet spacings sampled at the receiver to identify the capacity mode presented in their histogram (bprobe [1], PBM [25], Nettimer [17], and Pathrate [2]) or uses minimally delayed packet pairs to produce its estimate (CapProbe [16]). Another dimension of related work focuses on measuring available bandwidth of the tight link (Pathload [11], IMR-Pathload [15], IGI/PTR [6], and TOPP [23]). These techniques adjust sending rates of probe packets to infer the available bandwidth using relationship between the sending rate and corresponding receiving rate.…”