On Thu, Apr 11, 2024 at 1:19 PM Damien Le Moal <dlemoal@xxxxxxxxxx> wrote: > > On 4/10/24 22:17, Jens Axboe wrote: > > On 4/10/24 4:18 AM, Dongliang Cui wrote: > >> The default configuration in the current code is that when the device > >> is not busy, a single dispatch will attempt to pull 'nr_requests' > >> requests out of the schedule queue. > >> > >> I tried to track the dispatch process: > >> > >> COMM TYPE SEC_START IOPRIO INDEX > >> fio-17304 R 196798040 0x2005 0 > >> fio-17306 R 197060504 0x2005 1 > >> fio-17307 R 197346904 0x2005 2 > >> fio-17308 R 197609400 0x2005 3 > >> fio-17309 R 197873048 0x2005 4 > >> fio-17310 R 198134936 0x2005 5 > >> ... > >> fio-17237 R 197122936 0x0 57 > >> fio-17238 R 197384984 0x0 58 > >> <...>-17239 R 197647128 0x0 59 > >> fio-17240 R 197909208 0x0 60 > >> fio-17241 R 198171320 0x0 61 > >> fio-17242 R 198433432 0x0 62 > >> fio-17300 R 195744088 0x2005 0 > >> fio-17301 R 196008504 0x2005 0 > >> > >> The above data is calculated based on the block event trace, with each > >> column containing: process name, request type, sector start address, > >> IO priority. > >> > >> The INDEX represents the order in which the requests are extracted from > >> the scheduler queue during a single dispatch process. > >> > >> Some low-speed devices cannot process these requests at once, and they will > >> be requeued to hctx->dispatch and wait for the next issuance. > >> > >> There will be a problem here, when the IO priority is enabled, if you try > >> to dispatch "nr_request" requests at once, the IO priority will be ignored > >> from the scheduler queue and all requests will be extracted. > >> > >> In this scenario, if a high priority request is inserted into the scheduler > >> queue, it needs to wait for the low priority request in the hctx->dispatch > >> to be processed first. > >> > >> --------------------dispatch 1st---------------------- > >> fio-17241 R 198171320 0x0 61 > >> fio-17242 R 198433432 0x0 62 > >> --------------------dispatch 2nd---------------------- > >> fio-17300 R 195744088 0x2005 0 > >> > >> In certain scenarios, we hope that requests can be processed in order of io > >> priority as much as possible. > >> > >> Maybe max_dispatch should not be a fixed value, but can be adjusted > >> according to device conditions. > >> > >> So we give a interface to control the maximum value of single dispatch > >> so that users can configure it according to devices characteristics. > > > > I agree that pulling 'nr_requests' out of the scheduler will kind of > > defeat the purpose of the scheduler to some extent. But rather than add > > another knob that nobody knows about or ever will touch (and extra queue > > variables that just take up space), why not just default to something a > > bit saner? Eg we could default to 1/8 or 1/4 of the scheduler depth > > instead. Current mechanism will try to pulling everything out of the scheduler, regardless of the priority of request. Perhaps reducing the queue depth does not solve the priority disorder scenario. Reducing depth may also weakens the role of priorify. > > Why not default to pulling what can actually be executed, that is, up to the > number of free hw tags / budget ? Anything more than that will be requeued anyway. The process of pulling the request out of schedule will try to obtain the hw tags. If hw tags are obtained, request will continue to be pulled out of scheduler. However, for slow devices, the number of hw tags is generally greater than the number of requests that the device can currently handle, and unprocessed requests may not be arranged in dispatch_list in order of priority. And for budget, I found that the budget is set by the driver. Some slow devices, such as emmc, do not register this interface, unable to set budget. > -- > Damien Le Moal > Western Digital Research >
