File-access characteristics of parallel scientific workloads

Abstract

Phenomenal improvements in the computational performance of multiprocessors have not been matched by comparable gains in I/O system performance. This imbalance has resulted in I/O becoming a signi cant bottleneck for many scienti c applications. One key to overcoming this bottleneck is improving the performance of parallel le systems. The design of a high-performance parallel le system requires a comprehensive understanding of the expected workload. Unfortunately, u n til recently, no general workload studies of parallel le systems have been conducted. The goal of the CHARISMA project was to remedy this problem by c haracterizing the behavior of several production workloads, on di erent machines, at the level of individual reads and writes. The rst set of results from the CHARISMA project describe the workloads observed on an Intel iPSC/860 and a Thinking Machines CM-5. This paper is intended to compare and contrast these two w orkloads for an understanding of their essential similarities and di erences, isolating common trends and platform-dependent v ariances. Using this comparison, we are able to gain more insight i n to the general principles that should guide parallel le-system design.

Workload Modeling for Computer Systems Performance Evaluation

“…Table 9.4. ent processes can be finely interleaved in the file [415,516]. This is reflected in the file access modes of Table 9.4 and Figure 9.45, which typically reflect a partitioning of some data among the processors.…”

“…In a study from the mid-1990s, most I/O operations were in the range of tens to hundreds of bytes long, but most of the data being read or written were part of the few very large operations typically involving (many) megabytes [516]. This pattern leads to significant mass-count disparity.…”

“…Therefore different accesses interleave with each other; in many cases, this is an interleav-ing of streams of strided access used to read or write multidimensional data structures [149,515,415,516,554]. This interleaving leads to a phenomenon known as interprocess locality, because accesses by one process allow the system to predict accesses by other processes [412].…”

“…At about the same time, in work on parallel I/O, I was exposed to the Charisma I/O traces collected by David Kotz and Nils Nieuwejaar [516]. Among the voluminous data on I/O operations were a few records about the jobs to which they belonged.…”

Automated Tracing of I/O Stack

“…However, their strategy does not provide flexibility in terms of target metric specification. Tools such as CHARISMA [20], Pablo [23], and TAU (Tuning and Analysis Utilities) [19] are designed to collect and analyze file system traces [18]. For the MPI-based parallel applications, several tools, such as MPE (MPI Parallel Environment) [4] and mpiP [26], exist.…”