Hi, > On Sun, May 21, 2023 at 08:49:53AM +0800, Wang Yugui wrote: > > Hi, > > > > > Wang Yugui has a workload which would be improved by using large folios. > > > Until now, we've only created large folios in the readahead path, > > > but this workload writes without reading. The decision of what size > > > folio to create is based purely on the size of the write() call (unlike > > > readahead where we keep history and can choose to create larger folios > > > based on that history even if individual reads are small). > > > > > > The third patch looks like it's an optional extra but is actually needed > > > for the first two patches to work in the write path, otherwise it limits > > > the length that iomap_get_folio() sees to PAGE_SIZE. > > > > very good test result on 6.4.0-rc2. > > # just drop ';' in 'if (bytes > folio_size(folio) - offset);' of [PATCH 3/3]. > > > > fio -name write-bandwidth -rw=write -bs=1024Ki -size=32Gi -runtime=30 -iodepth 1 > > -ioengine sync -zero_buffers=1 -direct=0 -end_fsync=1 -numjobs=4 > > -directory=/mnt/test > > fio WRITE: bw=7655MiB/s (8027MB/s). > > > > Now it is the same as 5.15.y > > Great! How close is that to saturating the theoretical write bandwidth > of your storage? SSD: PCIe3 SSD U.2 *4 (1.6T/3.2T/3.2T/3.2T, just 800G used) through a NVMe adapater. CPU: E5-2680 v2 @ 2.80GHz *2 memory: DDR3 * 3 channel so theoretical write bandwidth maybe over 12GB/s, but 8GB/s maybe the in fact bandwidth because of - write depth limit - NVMe adapther limit - CPU/memory limit I also noticed a huge improvement for single thread dd. # 6.4.0-rc2 with this patch #dd-9.1 conv=fsync bs=1024K count=32K if=/dev/zero of=/mnt/test/dd.txt 34359738368 bytes (34 GB, 32 GiB) copied, 6.96108 s, 4.6 GiB/s But it is about 2.2GiB/s in xfs/5.15.y. Best Regards Wang Yugui (wangyugui@xxxxxxxxxxxx) 2023/05/21
