On 5/30/24 3:17 PM, Keith Busch wrote: > On Thu, May 30, 2024 at 02:02:20PM -0700, Bart Van Assche wrote: >> Thank you for having run this test. I propose that users who want better >> fairness than what my patch supports use an appropriate mechanism for >> improving fairness (e.g. blk-iocost or blk-iolat). This leaves the choice >> between maximum performance and maximum fairness to the user. Does this >> sound good to you? > > I really don't know, I generally test with low latency devices and > disable those blk services because their overhead is too high. I'm > probably not the target demographic for those mechanisms. :) Yeah same. But outside of that, needing to configure something else is also a bit of a cop out. From the initial posting, it's quoting 2.9% gain. For lots of cases, adding blk-iocost or blk-iolat would be MORE than a 2.9% hit. That said, I'd love to kill the code, but I still don't think we have good numbers on it. Are yours fully stable? What does the qd=1 test do _without_ having anyone compete with it? Is the bandwidth nicely balanced if each does qd=32? I'm again kindly asking for some testing :-) > I just wanted to push the edge cases to see where things diverge. > Perhaps Jens can weigh in on the impact and suggested remedies? Don't think we have enough data yet to make the call... -- Jens Axboe
