Need the configuration file of HSE

[ZhangJiaQiao]

Description

I want to run the YCSB test on HSE and compare it with RocksDB. In order to get the best performance of both engines for a fair comparision, I want to know the configuration of HSE and RocksDB for this YCSB test of Micron, or other way, some configuration advice.

[ZhangJiaQiao]

Using the default configuration of HSE and RocksDB, I can such YCSB test results as listed below.

RocksDB:

./bin/ycsb load rocksdb -P workloads/my_workload -p rocksdb.dir=./ycsb_data/ycsb-rocksdb-data -threads 30

[OVERALL], RunTime(ms), 1926
[OVERALL], Throughput(ops/sec), 51921.07995846314
[TOTAL_GCS_G1_Young_Generation], Count, 2
[TOTAL_GC_TIME_G1_Young_Generation], Time(ms), 16
[TOTAL_GC_TIME_%_G1_Young_Generation], Time(%), 0.8307372793354102
[TOTAL_GCS_G1_Old_Generation], Count, 0
[TOTAL_GC_TIME_G1_Old_Generation], Time(ms), 0
[TOTAL_GC_TIME_%_G1_Old_Generation], Time(%), 0.0
[TOTAL_GCs], Count, 2
[TOTAL_GC_TIME], Time(ms), 16
[TOTAL_GC_TIME_%], Time(%), 0.8307372793354102
[CLEANUP], Operations, 30
[CLEANUP], AverageLatency(us), 6920.9
[CLEANUP], MinLatency(us), 3
[CLEANUP], MaxLatency(us), 207487
[CLEANUP], 95thPercentileLatency(us), 29
[CLEANUP], 99thPercentileLatency(us), 207487
[INSERT], Operations, 100000
[INSERT], AverageLatency(us), 371.62749
[INSERT], MinLatency(us), 10
[INSERT], MaxLatency(us), 25295
[INSERT], 95thPercentileLatency(us), 626
[INSERT], 99thPercentileLatency(us), 1522
[INSERT], Return=OK, 100000

HSE:

LD_LIBRARY_PATH=/opt/hse/lib/x86_64-linux-gnu ./bin/ycsb load hse -P workloads/my_workload -p hse.kvdb.home=./ycsb_data/ycsb-hse-data -threads 30

[OVERALL], RunTime(ms), 7874
[OVERALL], Throughput(ops/sec), 12700.0254000508
[TOTAL_GCS_G1_Young_Generation], Count, 2
[TOTAL_GC_TIME_G1_Young_Generation], Time(ms), 20
[TOTAL_GC_TIME_%_G1_Young_Generation], Time(%), 0.25400050800101603
[TOTAL_GCS_G1_Old_Generation], Count, 0
[TOTAL_GC_TIME_G1_Old_Generation], Time(ms), 0
[TOTAL_GC_TIME_%_G1_Old_Generation], Time(%), 0.0
[TOTAL_GCs], Count, 2
[TOTAL_GC_TIME], Time(ms), 20
[TOTAL_GC_TIME_%], Time(%), 0.25400050800101603
[CLEANUP], Operations, 30
[CLEANUP], AverageLatency(us), 56171.6
[CLEANUP], MinLatency(us), 3
[CLEANUP], MaxLatency(us), 1685503
[CLEANUP], 95thPercentileLatency(us), 21
[CLEANUP], 99thPercentileLatency(us), 1685503
[INSERT], Operations, 100000
[INSERT], AverageLatency(us), 354.14383
[INSERT], MinLatency(us), 5
[INSERT], MaxLatency(us), 19599
[INSERT], 95thPercentileLatency(us), 1029
[INSERT], 99thPercentileLatency(us), 1108
[INSERT], Return=OK, 100000

The result I get shows that HSE get a worse performance than RocksDB, which does match what the Micron report declare.

[alexttx]

I suspect you're seeing the effect of HSE's slow start mechanism. At startup HSE throttles the input rates until it learns the capabilities of the underlying storage. This means that short benchmarks will not see HSE's full rate. See initial throttle setting for more information.

Also Note that many configuration parameters you see in source code are not meant to be changed except for debugging and development. We mark those parameters as "experimental". The non-experimental parameters are described here.

My suggestions are:

• run a longer benchmark. you should see the rate increase after a few minutes.
• set the initial throttle setting to "light" as described in the above link
• experiment with the thread count. we often run YCSB thread count between 1 and 1.5 times the number reported by nproc on systems with 128G of DRAM or more.

[smoyerx]

I also recommend that you review the benchmarking tips in the documentation.