I was testing 0.98.4-SNAPSHOT with YCSB on a small testbed for HBASE-11297 and took a small detour to try out collection and processing of Brendan Gregg's Java flame graphs ( http://www.brendangregg.com/blog/2014-06-12/java-flame-graphs.html) They seemed to turn out well. I want to look at them more closely next week.
For your viewing enjoyment. Traces were collected from the usual experiment, a 5 slave cluster with 5 concurrent YCSB clients generating a target aggregate load of 100,000 ops/second. Each RegionServer has 8 GB of heap and bucket cache enabled backed by 24 GB of off heap memory. The test table schema is all defaults except I turned on FAST_DIFF block encoding because I was also curious about its CPU usage profile (along with that of the bucket cache). One RegionServer was instrumented with Jeremy Manson's lightweight accurate Java profiler, with the kMaxStackTraces and kMaxFramesToCapture options in src/globals.h changed to 100000 and 128, respectively, out of pessimism and expectations for the running time of the experiment. Workload A: https://db.tt/oTqTmGIE Workload B: https://db.tt/jy3E5fh3 Workload C: https://db.tt/yFxgKROq Workload D: https://db.tt/36Ux659r Workload E: https://db.tt/Or2UuTKi Workload F: https://db.tt/4Wud5il4 For more information on flame graphs and how to read them, see http://www.slideshare.net/brendangregg/blazing-performance-with-flame-graphs -- Best regards, - Andy Problems worthy of attack prove their worth by hitting back. - Piet Hein (via Tom White)
