Hi,
在 2024/01/16 10:59, Bart Van Assche 写道:
On 1/14/24 22:18, Yu Kuai wrote:
Can you take a look at my previous patset if you haven't? And it'll
be great to hear from your comments.
https://lore.kernel.org/all/20231021154806.4019417-1-yukuai1@xxxxxxxxxxxxxxx/
Something is missing from the cover letter of that patch series:
measurements that show the impact of that patch series on the maximum
IOPS that can be achieved with the null_blk driver. I'm afraid that the
performance impact of that patch series will be larger than what is
acceptable.
Hi,
This version is just RFC, and is focusing on the method. I ran some
tests on null_blk in my VM and didn't notice performace regression. My
idea is that I already make this patchset complicated, and I'm looking
for some comments for this patchset without spending too much time on
this blindly.
I'll provide null_blk tests result in the next version if anyone thinks
the approch is acceptable:
1) add a new field 'available_tags' and update it in slow path, hence
fast path hctx_may_queue() won't be affected.
2) delay tag sharing untill failed to get driver tag;
3) add a timer per shared tags to balance shared tags;
Thanks,
Kuai
Thanks,
Bart.
.
