Closed Jerry-Tianchen closed 11 months ago
You're right, there should be a hdr_reset()
before changing the message size. But line 195 might not be the right place for this, since we also want to track longer-term percentiles beyond one measurement interval. So it needs to be coupled with setting of double_req_size_
.
There are a few other boundary issues with the measurement (e.g., req size increases when some smaller old requests are in-flight) too. I've mostly been using this benchmark with the runtime msg size changes disabled.
I see~ I will also take a look into the issue you say. update you later xD
Thanks Jerry
Hi"
I realize that there is an error in the calculation of the latency. Apparently, you did not reset the HDR when trying different block sizes....
It will result in large block sizes also have latency like 2.7 us. Please add
hdr_reset(c.latency_hist_);
under line 195 for app/latency.cc