Dynamic Code Management

Poor code locality degrades application performance, especially for large programs such as server applications, by increasing the memory stalls caused by instruction TLB and instruction cache misses. This locality problem is a particular issue for languages such as Java and C# that provide just-in-time (JIT) compilation, dynamic class loading, and dynamic recompilation. Managed runtimes for these languages offer the opportunity to dynamically profile an application’s execution and adapt that execution to improve performance. This paper describes the dynamic code management system (DCM) of our managed runtime that uses dynamic profile information to reorder JIT-compiled code when necessary to improve locality. To place method code, our DCM system supports the widely-used Pettis-Hansen procedure placement algorithm. In addition to this algorithm, we also developed and implemented three new placement algorithms. Our new algorithms can run up to 6000 times faster than the Pettis-Henson algorithm and still achieve the same or better performance improvement, which makes them more suitable for use in high-performance managed runtimes. The new algorithms specifically target ITLB misses, which typically have the greatest impact on performance. Using the DCM, we demonstrate a 6% performance improvement with our new Graph-Walking layout algorithm for the large MiniBean benchmark, a single-process version of the SPECjAppServer20021 enterprise application server benchmark. This performance is 4% better than that provided by Pettis-Hansen placement. Furthermore, instead of the prohibitively expensive 35 minutes Pettis-Hansen layout requires, Graph-Walking layout needs just 0.35 second (6000 times faster). ∗This work is supported by NSF ITR CCR-0085792, NSF CCR0311829, NSF EIA-0303609, DARPA F33615-03-C-4106, ARC DP0452011 and Intel. Any opinions, findings and conclusions expressed herein are the authors and do not necessarily reflect those of the sponsors. 1We use the SPEC benchmarks only as benchmarks to compare the performance of the various techniques within our own VM. We are not using them to compare our VM to any other VM and are not publishing SPEC metrics of any kind. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. VEE’06, June 14-16, 2006, Ottawa, Canada. Copyright 2006 ACM ...$5.00.