F. T. Chong, R. Barua, F. Dahlgren, J. D. Kubiatowicz, and A. Aragwal, "The Sensitivity of Communication Mechanisms to Bandwidth and Latency", Proceedings of the 4th Int'l Symposium on High Performance Computer Architecture, Las Vegas, NV, Feb 1-4, 1998.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....in the processor. The message passing programs have much coarser grained communications than shared memory programs, millions of instructions between communication operations rather than hundreds of instructions, making it feasible for ULTRA to store traces. An approach taken by Chong, et al. [Chong98] to make the data collection computer s characteristics more closely match the architecture being studied was to use a portion of processors in the machine to produce additional traffic in the data collection computer to adjust the effective bandwidth and latency. To 5 10:04:16 October 1, 1999 ....

F. T. Chong, R. Barua, F. Dahlgren, J. D. Kubiatowicz, and A. Aragwal, "The Sensitivity of Communication Mechanisms to Bandwidth and Latency", Proceedings of the 4th Int'l Symposium on High Performance Computer Architecture, Las Vegas, NV, Feb 1-4, 1998.

....synchronization, such as exclusive locks. Some of these applications perform well with the base polling model because they communicate frequently. 3. SIMULATION ENVIRONMENT The simulator used in this study is SimCVM, an execution based simulator that models the CVM (Coherent Virtual Machine) [7] software DSM. CVM is a user level library that features a set of base classes that provide a framework for implementing specific DSM protocols. These classes include a generic protocol class, a 0 10 20 30 40 50 60 70 barnes fft sor spatial water time blocked on remote ....

....do, but also at the hardware level for masking cache misses [17] In this work, we show how multi threading compares in hiding the remote request latencies vs. fast software interrupts in reducing those latencies. The sensitivity of communication mechanisms to message latency was studied by Chong [7]. The main issue studied here is how shared memory and message passing compare to each other with respect to network latency. For message passing, they agree with Brewer [6] that polling performs better when performance is dominated by communication. In addition to direct costs, interrupts can ....

[Article contains additional citation context not shown here]

F. Chong, R. Barua, F. Dahlgren, J. Kubiatowicz, and A. Agarwal, "Sensitivity of Communication Mechanisms to Bandwidth and Latency," in Proceedings of 4 th Int'l Symposium on High Performance Computer Architecture (HPCA), February 1998.

....communication to make request distribution decisions that are as accurate as possible. In fact, as far as we know, PRESS is the first server to apply user level communication and in particular the VIA standard. User level communication has been studied in several previous papers (e.g. [15, 8, 22, 18]) However, all of these studies were performed in the context of scientific applications, microbenchmarks, or distributed shared memory systems. Our study focuses on an important non scientific, I Ointensive application, network servers. In addition, although a detailed study of VIA is not our ....

F. Chong, R. Barua, F. Dahlgren, J. Kubiatowicz, and A. Agarwal. The Sensitivity of Communication Mechanisms to Bandwidth and Latency. In Proceedings of the 4th International Symposium on High-Performance Computer Architecture, January 1998.

....to make request distribution decisions that are as accurate as possible. In fact, as far as we know, PRESS is the first cluster based server to apply user level communication and in particular the VIA standard. User level communication has been studied in several previous papers (e.g. [19, 9, 27, 22]) However, all of these studies were performed in the context of scientific applications or distributed sharedmemory systems. Our study focuses on an important non scientific application, network servers. In addition, although a detailed study of VIA is not our focus, this paper does extend the ....

F. Chong, R. Barua, F. Dahlgren, J. Kubiatowicz, and A. Agarwal. The Sensitivity of Communication Mechanisms to Bandwidth and Latency. In Proceedings of the 4th International Symposium on High-Performance Computer Architecture, January 1998.

....improves. In the other two cases, update performs worse than invalidate. This is consistent with the results reported earlier in this chapter update performance varies widely with network parameters and is especially sensitive to changes in network bandwidth. 5.7. 3 Message Passing Chong et al. [19] compare the performance of shared memory and message passing on the Alewife system. Unlike our study, their experiments were performed on actual Alewife hardware. The benchmarks are three engineering codes featuring irregular meshes or graphs (EM3D, UNSTRUC, and ICCG) and one molecular dynamics ....

Frederic T. Chong, Rajeev Barua, Fredrik Dahlgren, John D. Kubiatowicz, and Anant Agarwal. The sensitivity of communication mechanisms to bandwidth and latency. In Fourth International Symposium on High Performance Computer Architecture, February 1998.

....and discusses implications of this work. 2 Related Work and Background In this paper, we try to understand how large differences in bandwidths and latencies in an interconnect influence application performance. For small gaps, several studies report good performance on hardware DSM NUMA systems [10, 19, 24, 33, 34, 41]. These systems have a gap of about a factor of 3 between the slowest and the fastest links. The picture changes for systems with longer access times. Papers by [11, 22, 25, 31, 36, 42] study local area clusters of SMPs, which have a gap of up to an order of magnitude. These studies tend to focus ....

....studies use small, fixed, gaps. We are interested in how performance scales with different gaps; in our interconnect the gap is varied over a large range, from zero to four orders of magnitude. In addition, many studies focus on issues such as DSM protocols or message passing versus shared memory [9, 10, 35]. Again, our focus is the NUMA gap. We investigate where conventional applications break down, how communication patterns can be adapted, and how far performance improvement can be pushed. As an important aside, we want to know how difficult it is to implement such changes. Differences in link ....

F. Chong, R. Barua, F. Dahlgren, J. Kubiatowicz, and A. Agarwal. The Sensitivity of Communication Mechanisms to Bandwidth and Lantency. In HPCA-4 High Performance Communication Architectures, pages 37--46, February 1998.

....on distributed memory machines. In contrast, our benchmark programs are mostly written as coarse grained message passing programs. Recent work by Chong et al. on the Alewife machine shows that message passing programs are relatively insensitive to latency in comparison to shared memoryprograms [5]. This is a consequence of message passing programs exploiting asynchronous communications to tolerate network latency by overlapping communication and computation. 9. Conclusions We investigated the feasibility of clusters of workstations to run challenging parallel applications. Since the ....

F. Chong, R. Barua, F. Dahlgren, J. Kubiatowicz, and A. Agarwal. The Sensitivity of Communication Mechanisms to Bandwidth and Latency. In HPCA-4, Las Vegas, Nevada, Feb. 1998.

....arising from the integration of four features desired in a scalable workstation. Message passing and shared memory are features desired for high performance parallel processing while multiprogramming and virtual memory are features desired for general purpose, multiuser operation. tolerance [12]. In particular, even if written assuming a shared memory programming model, a program in which communication patterns are amenable to compile time analysis might make better use of compiler generated message passing communication than of even hardware supported shared memory [14, 12] Third, ....

.... tolerance [12] In particular, even if written assuming a shared memory programming model, a program in which communication patterns are amenable to compile time analysis might make better use of compiler generated message passing communication than of even hardware supported shared memory [14, 12]. Third, distributed shared memory and message passing implementations are naturally similar at a low level, so if an implementation provides shared memory in hardware the additional cost of exposing message passing is low [40] Finally, one reading of technology trends is that shared memory ....

Frederic T. Chong, Rajeev Barua, Fredrik Dahlgren, John D. Kubiatowicz, and Anant Agarwal. The Sensitivity of Communication Mechanisms to Bandwidth and Latency. In Proceedings of the Fourth International Symposium on High-Performance Computer Architecture (HPCA-4), February 1998.

....waves in solids. Our validations are performed against the MIT Alewife multiprocessor which was parameterized in Section 2. Network bandwidth (and therefore contention) on Alewife is fairly representative of what is found in a variety of current generation commercial and research multiprocessors [7]. Because processor speeds are increasing faster than the bandwidth of high end local area networks (e.g. Myrinet) network contention effects on networks of workstations (e.g. 20] will be somewhat more severe than the contention observed in these results. 4.1 All to all remap In this subsection ....

....cache coherent distributed shared memory machines using four parameters similar to LogP. They used the o performance parameter to model the occupancy of the communication controller. Their study shows that application performance is highly sensitive to the controller occupancy. Finally Chong et al. [7] examined the effect of network bandwidth on application performance using several different communication schemes. They find that message passing communication primitives are less sensitive to network bandwidth than are shared memory primitives. Several studies apply the contention free LogP ....

F. Chong, R. Barua, F. Dahlgren, J. Kubiatowicz and A. Agarwal, "The Sensitivity of Communication Mechanisms to Bandwidth and Latency," Proc. of the 4th Int'l Symp. on High Performance Computer Architecture, Las Vegas, NV, Feb. 1998.

....and discusses implications of this work. 2 Related Work and Background In this paper, we try to understand how large differences in bandwidths and latencies in an interconnect influence application performance. For small gaps, several studies report good performance on hardware DSM NUMA systems [9, 17, 21, 29, 35, 28]. These systems have a gap of about a factor of 3 between the slowest and the fastest links. The picture changes for systems with longer access times. Papers by [10, 19, 22, 26, 31, 36] study local area clusters of SMPs, which have a gap of up to an order of magnitude. These studies tend to focus ....

....studies use small, fixed, gaps. We are interested in how performance scales with different gaps; in our interconnect the gap is varied over a large range, from zero to four orders of magnitude. In addition, many studies focus on issues such as DSM protocols or message passing versus shared memory [8, 9, 30]. Again, our focus is the NUMA gap. We investigate where conventional applications break down, how communication patterns can be adapted, and how far performance improvement can be pushed. As an important aside, we want to know how difficult it is to implement such changes. Differences in link ....

F. Chong, R. Barua, F. Dahlgren, J. Kubiatowicz, and A. Agarwal. The Sensitivity of Communication Mechanisms to Bandwidth and Lantency. In HPCA-4 High Performance Communication Architectures, pages 37--46, February 1998.

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