Lossy Compression of Communication Traces Using Recurrent Neural Networks

Sun, Jingwei; Yan, Tao; Sun, Hao; Lin, Huancheng; Sun, Guoqiang

doi:10.1109/tpds.2021.3132417

Cited by 6 publications

(7 citation statements)

References 38 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This definition is based on two points. (1) Existing studies [13,15,24,31,34,43] prefer to use the interval between two communication events to indicate a portion of computation time cost. Our definition aligns with the convention.…”

Section: Tracing Computation Eventsmentioning

confidence: 99%

“…The major difficulty is the huge storage consumption of behavior record, which is often called trace, from multiple cluster nodes. For instance, tracing a single execution of a mini-app LULESH [18] on less than 1,000 processors can produce hundreds of gigabytes trace data [31]. As a compromise, existing methods mainly focus on communication trace [13,15,24,31,34,43], since communication events constitute the skeleton of an HPC application.…”

Section: Introductionmentioning

confidence: 99%

“…For instance, tracing a single execution of a mini-app LULESH [18] on less than 1,000 processors can produce hundreds of gigabytes trace data [31]. As a compromise, existing methods mainly focus on communication trace [13,15,24,31,34,43], since communication events constitute the skeleton of an HPC application. Meanwhile, computation behaviors are often overlooked.…”

Section: Introductionmentioning

confidence: 99%

“…Meanwhile, computation behaviors are often overlooked. Computation behaviors are more diverse and difficult to trace, so they are usually recorded as time intervals and replayed by calling sleep functions or empty loops [13,15,24,31,34,43]. Besides, lossy compression methods are used to further reduce the storage cost of communication trace, of which the original communication parameters and timing information are approximated by histogram [23,34,36], clustering [2,3], and stochastic process [8,15,31].…”

Section: Introductionmentioning

confidence: 99%

“…Computation behaviors are more diverse and difficult to trace, so they are usually recorded as time intervals and replayed by calling sleep functions or empty loops [13,15,24,31,34,43]. Besides, lossy compression methods are used to further reduce the storage cost of communication trace, of which the original communication parameters and timing information are approximated by histogram [23,34,36], clustering [2,3], and stochastic process [8,15,31]. Due to the ignorance of computation details and the lossy compression on communication trace, current synthetic proxy-apps are difficult to comprehensively resemble the performance characteristics of HPC applications.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Synthesizing Proxy Applications for MPI Programs

Yan¹,

Xu²,

Luo³

et al. 2023

Preprint

View full text Add to dashboard Cite

Proxy applications (proxy-apps) are fundamental tools for evaluating the performance of specific workloads on high-performance computing (HPC) systems. Since the development of high-fidelity proxy-apps, which exhibit similar performance characteristics as corresponding production applications, is labor-intensive, synthetic proxy-apps are created as a useful supplement to manually-developed proxy-apps. Existing proxyapp synthesis methods for HPC applications mainly focus on either communications patterns or computation kernels. To thoroughly resemble performance characteristics of HPC applications represented by Message Passing Interface (MPI) programs, we propose Siesta, a novel framework to automatically synthesize proxyapps based on communication-computation traces. Given an MPI program, Siesta synthesizes parameterized code proxies to mimic computation behaviors in different execution periods, and combines the code snippets and MPI function records into an event trace. It then extracts program behavior patterns from the trace using context-free grammars, and finally transforms the grammars into executable code as a synthetic proxy-app of the given MPI program. We evaluate the proposed methods on representative MPI programs with different parallel scales. The results show that our synthetic proxy-apps can precisely approximate the performance characteristics of MPI programs.

show abstract