VESPA: static profiling for binary optimization

Abstract

Over the past few years, there has been a surge in the popularity of binary optimizers such as BOLT, Propeller, Janus and HALO. These tools use dynamic profiling information to make optimization decisions. Although effective, gathering runtime data presents developers with inconveniences such as unrepresentative inputs, the need to accommodate software modifications, and longer build times. In this paper, we revisit the static profiling technique proposed by Calder et al. in the late 90’s, and investigate its application to drive binary optimizations, in the context of the BOLT binary optimizer, as a replacement for dynamic profiling. A few core modifications to Calder et al.’s original proposal, consisting of new program features and a new regression model, are sufficient to enable some of the gains obtained through runtime profiling. An evaluation of BOLT powered by our static profiler on four large benchmarks (clang, GCC, MySQL and PostgreSQL) yields binaries that are 5.47 % faster than the executables produced by clang -O3.

Beetle: A Feature-Based Approach to Reduce Staleness in Profile Data

“…In total, our feature vectors contain 36 dimensions. These dimensions were filtered from an initial collection of 63 features, out of which 21 were proposed by Calder et al [1997], and 35 were proposed by Moreira et al [2021]. The remaining 7 features were readily available in BOLT's implementation.…”

“…This general approach was called evidence-based static profiling by Calder et al [1997]. Although not new, recent developments in machine learning have returned evidence-based methodologies to the spotlight Moreira et al [2021], Sweety and Chaudhary [2021], Ugur et al [2022].…”

“…Key to these gains is dynamic profile information-data sampled during the execution of the target program. To put these numbers in perspective, recently Moreira et al [2021] attempted to replace BOLT's dynamic profiler with a state-of-the-art stochastic model that infers profile data statically. They improved the performance of Clang's binary by 8.80% -a lackluster result compared with BOLT's original gain of 37.…”

Thermometer

“…Prior work from Google and Facebook shows that their widely-deployed data center applications lose more than 15% of all pipeline slots due to frontend stalls [25,27,67,133]. As these applications are proprietary, we use the applications used by prior work [75,77,78,86,100,138,150], where frontend stalls are similarly frequent (more than 15%) due to large instruction footprints. These applications include cassandra [2], kafka [3], and tomcat [4] from the Java DaCapo benchmark suite [31], drupal [142], wordpress [144], and mediawiki [143] from Facebook's OSS -performance benchmark suite [16], finagle-chirper and finagle-http [12] from the Java Renaissance benchmark suite [114], clang [6] while building LLVM [85], PostgreSQL [10] while serving pgbench [9] queries, Python [14] while running the pyperformance [11] benchmark suite, MySQL [146] while serving TPC-C queries [35], and verilator [13] while emulating the Rocket Chip [7].…”

“…Hence, there is a significant performance gap between the state-of-the-art and the optimal replacement policy. code layout optimization techniques are sensitive to profile quality [55], they work exceptionally well in practice [27,33,47,100,106,107]. Profiles for data center applications change slowly over several weeks [33] while data center operators profile and recompile applications multiple times a day [21,33,106,107].…”