Performance of s with Memory Virtualization using IBM Active Memory Expansion as an example 5th International Workshop on Virtualization Technologies in Distributed Computing (VTDC) Marcus Homann Technical University Munich
Agenda Performance of s: Research at Technical University Munich Background & Motivation Performance Measurement Process Performance Measurements Results Conclusion and next Steps 2
Performance of s: Research at Technical University Munich Stephan Gradl: Performance simulation with increasing number of concurrent users Focus on ABAP-Stack Andre Bögelsack: Critical load Focus on ABAP-Stack Comparing several virtual machines Marcus Homann: Critical load Focus on ABAP-Stack Focus on mainmemory-compression ABAP Performance Simulation Virtualization Main-Memory Compression J2EE Performance Measurement Manuel Mayer: Performance simulation with increasing number of concurrent users Focus on Portal (J2EE-Stack) Holger Jehle: Average load Focus on J2EE-Stack Investigation of 1 virtual machine 3
In one sentence How does main-memory virtualization affect the performance of systems and which recommendations can be derived for data center operations? 4
Background & Motivation (1) Scenario 1: Without Main-Memory Compression Scenario 2: With Main-Memory Compression Virtual Main Memory Physical Main Memory Main memory compression Physical Main Memory 5
Background & Motivation (1) Scenario 1: Without Main-Memory Compression Scenario 2: With Main-Memory Compression Virtual Main Memory Physical Main Memory Main memory compression Physical Main Memory Performance? 6
Background & Motivation (2) Main-memory compression expands the main-memory capacity, but can negatively affect the application performance Concept: Main-Memory Compression Performance of Main-Memory Compression Application Throughput Physical Main- Memory Compression Uncompressed main-memory data Compressed main-memory data Application Response Time CPU Utilization Main-Memory Expansion Factor (Michel 2010, p. 5) (Michel 2010, p. 7) 7
Assumptions and Research Questions A1 A2 RQ1 RQ2 RQ3 The performance of systems is influenced negatively at a certain main-memory expansion factor. Using main-memory compression, additional systems can be operated on a physical server without any performance degradation. Which main-memory compression techniques exist in literature, how is their performance evaluated and which performance results are available specific for based workloads? To what extent do different main-memory expansion factors affect the performance of systems? Which recommendations can be given based on the performance measurement results of RQ2? 8
LitReview: Performance of main-memory virtualization Literature review shows that there is little knowledge about the performance behavior of systems using main-memory virtualization. Main-memory compression is no new topic (Douglis 1993, Kaplan 1999) Distinction between hardware- and software-based main-memory compression techniques; there is a trend towards software-based techniques Only recently available in products of major virtualization vendors Evaluation is mainly based on the hardware-oriented SPEC CPU benchmark suite Only one paper can be found where a workload is used for performance evaluation (Michel 2010); however the author does not describe what load generator he uses and how his test environment looks like. An detailed study about the performance behavior of systems using main-memory compression is missing 9
Performance Measurement Process Environment: IBM Power 750 Server (512 GB RAM, 4 CPUs, 32 Cores, 3,3 GHz) LPAR: 4 virtual processors, 0.1 processing unit each) ECC system EHP 4 (64 configured workprocesses) Load Generator and Measurement Tool: Zachmanntest (Bögelsack et. al 2011) Synthetic benchmark, simulates a power user Uses internal tables of the application server Outcome: throughput of the environment in rows per second 2 general Test setups: native, AME Variables: Number of parallel Zachmanntests (~ generated Load): 1, 2, 3, 6, 14, 20, 164) AME factor: 1.0, 1.3, 3.0, 5.0, 10.0 Values of interest: Throughput (Zachmanntest: rows per second) Three runs per test setting: result is arithmetic mean 10
Measurement Results 11
Conclusion and next Steps 1. The performance of a system is influenced by activating AME. 2. At some point during the execution, a system may encounter a huge performance collapse. This is especially true when choosing a very high AME memory expansion factor, e.g. 5.0, 10.0. 3. The performance of a system is influenced by both the activation of AME and the work load. 4. At peak performance the AME factor seem to have no influence 5. Our proposed baseline with AME=1.0 does not reflect the best performance. Instead, the best performance is reached with AME=1.3. Next Steps Gaining better understanding of AIX memory management Testing with a finer granuarity of AME steps 12
References Douglis, F.: The Compression Cache: Using On-line Compression to Extend Physical Memory. In: USENIX Conference, 1993, pp. 519-529. Kaplan, S. F.: Compressed Caching and Modern Virtual Memory Simulation. Disseration at University of Texas, Austin 1999. Hepkin, D.: Active Memory Expansion: Overview and Usage Guide. IBM Whitepaper 2010. Hevner, A.; Chatterjee, S.: Design Research in Information s. Springer Verlag, Berlin 2010. Michel, D.: Active Memory Expansion Performance. IBM Whitepaper, 2010. Tremaine, R. B., Franaszek, P. A., Robinson, J. T., Schulz, C. O., Smith, T. B., Wazlowski, M. E.; Bland, P. M.:IBM Memory Expansion Technology (MXT). IBM Journal of Research and Development, Vol. 45, No. 2, 2001, p. 271-285. Tuduce, I.C. and T. Gross: Adaptive main memory compression. USENIX Association, 2005. 13