Why we see difference in throughput although the average latency is not differing much? #1704
-
|
Avg latency: 1967 usec & Throughput: 6268 MiB/s Avg latency: 2249 usec & Throughput: 3039 MiB/s As we can see clearly, average latency differs only by 300 usec but throughput is almost 2x.
|
Beta Was this translation helpful? Give feedback.
Replies: 6 comments 7 replies
-
|
It is difficult to provide feedback without the full job file(s) used to produce the output above. |
Beta Was this translation helpful? Give feedback.
-
|
In general latency is the difference between the start time for an |
Beta Was this translation helpful? Give feedback.
-
|
Thanks @vincentkfu to look into this ! The above two outputs are generated by the same fio spec pasted above. The only difference is workload directory - one is SSD and one is gcsfuse mounted directory (fuse based system). To differentiate further b/w both the workload dirctory:
|
Beta Was this translation helpful? Give feedback.
-
|
The first quoted block contains a call to So I think we have an explanation for the discrepancy between latency and bandwidth. The latency values are similar but bandwidth differs because time required to open a file is not included in the latency measurement. For more details about latency measurement see: https://github.com/vincentkfu/fio-blog/wiki/How-does-fio-measure-latency |
Beta Was this translation helpful? Give feedback.
-
|
The clat/slat discrepancy between the two file systems is unlikely to be related to fio. Consider using |
Beta Was this translation helpful? Give feedback.
-
|
I think this discussion has run its course and the main question in the title was answered by Vincent. |
Beta Was this translation helpful? Give feedback.
io_u.c:get_io_u()has the information you are looking for: