benchmarking - YCSB on rocksdb: the latency becomes lower when the request issuing rate becomes higher

Question

I want to test the performance of mongodb. And I choose the workloadc, which is a read-only workload. So I load about 10G data. First, I did not set the 'target' parameter. And the test command is : ./bin/ycsb run mongodb -s -P workloads/workloadc the result is:

[READ], AverageLatency(us), 135.2952
[READ], MinLatency(us), 84
[READ], MaxLatency(us), 75455
[READ], 95thPercentileLatency(us), 196
[READ], 99thPercentileLatency(us), 271
[READ], Return=OK, 300000

Second, I set the 'target' parameter to 4000. The test command is: ./bin/ycsb run mongodb -s -P workloads/workloadc -target 4000 the result is:

[READ], AverageLatency(us), 221.37833987374168
[READ], MinLatency(us), 87
[READ], MaxLatency(us), 98239
[READ], 95thPercentileLatency(us), 403
[READ], 99thPercentileLatency(us), 459
[READ], Return=OK, 117220

The read latency of first experiment is 135us, but the read latency of second experiment is 221us. The latency becomes much higer when I try to set the 'target' to 1000 or smaller. Why the read latency becomes higher when I try to limit the throughput? I expect the latency becomes lower.

Note:I run these experiments on a machine with 32-core CPU and 32G DRAM. Although my data size is 10G, which can be all cached in the DRAM, I evit the data from the cache with command echo 1 > /proc/sys/vm/drop_caches before I run each expeiment.

Is there something wrong with my operation？Or is there something wrong with the time statistics in the source code?

Answer 1

等待大神答复