On 11/22/19 12:09 AM, Ming Lei wrote:
On Thu, Nov 21, 2019 at 04:45:48PM +0100, Hannes Reinecke wrote:
On 11/21/19 1:53 AM, Ming Lei wrote:
On Wed, Nov 20, 2019 at 11:05:24AM +0100, Hannes Reinecke wrote:
I would far prefer if we could delegate any queueing decision to the
elevators, and completely drop the device_busy flag for all devices.
If you drop it, you may create big sequential IO performance drop
on HDD., that is why this patch only bypasses sdev->queue_depth on
SSD. NVMe bypasses it because no one uses HDD. via NVMe.
I still wonder how much performance drop we actually see; what seems to
happen is that device_busy just arbitrary pushes back to the block
layer, giving it more time to do merging.
I do think we can do better then that...
For example, running the following script[1] on 4-core VM:
------------------------------------------
| QD:255 | QD: 32 |
------------------------------------------
fio read throughput | 825MB/s | 1432MB/s|
------------------------------------------
[ ... ]
Hi Ming,
Thanks for having shared these numbers. I think this is very useful
information. Do these results show the performance drop that happens if
/sys/block/.../device/queue_depth exceeds .can_queue? What I am
wondering about is how important these results are in the context of
this discussion. Are there any modern SCSI devices for which a SCSI LLD
sets scsi_host->can_queue and scsi_host->cmd_per_lun such that the
device responds with BUSY? What surprised me is that only three SCSI
LLDs call scsi_track_queue_full() (mptsas, bfa, esp_scsi). Does that
mean that BUSY responses from a SCSI device or HBA are rare?
Thanks,
Bart.
