CacheBench latency reporting

[SuhoSon]

I had a question about the latency values of the cachebench output.
Are NVM read/write latencies cumulative since startup? or specific period value?

== Allocator Stats ==
Items in RAM  : 17,577,609
Items in NVM  : 1,251,040,911
Alloc Attempts: 4,492,637,019 Success: 100.00%
RAM Evictions : 4,475,056,085
Fraction of pool 0 used : 0.82
Fraction of pool 1 used : 0.99
Cache Gets    : 7,198,360,206
Hit Ratio     :  80.81%
RAM Hit Ratio :  37.59%
NVM Hit Ratio :  69.25%
RAM eviction rejects expiry : 0
RAM eviction rejects clean : 3,095,289,386
NVM Read  Latency    p50      :      76.00 us
NVM Read  Latency    p90      :     140.00 us
NVM Read  Latency    p99      :     268.00 us
NVM Read  Latency    p999     :     798.00 us
NVM Read  Latency    p9999    :    3663.00 us
NVM Read  Latency    p99999   :    4600.00 us
NVM Read  Latency    p999999  :    4693.00 us
NVM Read  Latency    p100     :       0.00 us
NVM Write Latency    p50      :      21.00 us
NVM Write Latency    p90      :      49.00 us
NVM Write Latency    p99      :     154.00 us
NVM Write Latency    p999     :     358.00 us
NVM Write Latency    p9999    :    1396.00 us
NVM Write Latency    p99999   :    1798.00 us
NVM Write Latency    p999999  :    1838.00 us
NVM Write Latency    p100     :       0.00 us
NVM bytes written (physical)  : 5812.23 GB
NVM bytes written (logical)   : 1135.59 GB
NVM bytes written (nand)      : 6501.07 GB
NVM app write amplification   :   5.12
NVM dev write amplification   :   1.12
NVM Gets      :   4,492,636,732, Coalesced :   0.00%
NVM Puts      :   1,379,841,526, Success   : 100.00%, Clean   :   0.22%, AbortsFromDel   :       27, AbortsFromGet   :    1,171
NVM Evicts    :     128,799,621, Clean     :   2.46%, Unclean :       0, Double          :        0
NVM Deletes   :   1,381,286,441 Skipped Deletes:  98.94%
Released 36,940 slabs
  Moves     : attempts:          0, success: 100.00%
  Evictions : attempts: 199,841,363, success:  99.96%

[haowu14]

I assume these are printed when your run finishes. Then it is accounting for all the NVM read/writes the cache received ever since it started.

[SuhoSon]

Thank you for responding to my question.
I think it should do that too. However, these log messages are printed periodically for a minute.
So, I looked into the src code for analyze and found that "tracker" uses the "SlidingWindowQuantileEstimator" which accumulates latency data for the duration of "window".