On 27/01/26 04:16PM, Christoph Hellwig wrote:
On Tue, Jan 27, 2026 at 08:24:28PM +0530, Anuj gupta wrote:
Hi Christoph,
Here are the QD1 latency numbers (in usec)
Thanks a lot!
Adding in the baseline numbers, as I wanted to compare those:
Intel Optane:
Sequential read
| size | baseline | xfs-bounce | xfs-pi |
+------+----------+-----------+-----------+
| 4k | 7.18 | 13.62 | 7.20 |
| 64K | 36.40 | 99.66 | 34.16 |
| 1M | 206.38 | 258.88 | 306.23 |
+------+----------+------------+----------+
So for 4k and 64k reads we basically get back to the baseline.
The 1M numbers are puzzling, though. I wonder if we need to
add WQ_CPU_INTENSIVE or do some other tweaks to the XFS I/O
completion workqueue so that we don't overload the scheduler.
Tried adding WQ_CPU_INTENSIVE[1] but didn't see any change in the 1M
latency. Looks like this needs something beyond just workqueue tuning.
