On 08/14/2017 09:38 AM, Jens Axboe wrote: > On 08/14/2017 09:31 AM, Minchan Kim wrote: >>> Secondly, generally you don't have slow devices and fast devices >>> intermingled when running workloads. That's the rare case. >> >> Not true. zRam is really popular swap for embedded devices where >> one of low cost product has a really poor slow nand compared to >> lz4/lzo [de]comression. > > I guess that's true for some cases. But as I said earlier, the recycling > really doesn't care about this at all. They can happily coexist, and not > step on each others toes. Dusted it off, result is here against -rc5: http://git.kernel.dk/cgit/linux-block/log/?h=cpu-alloc-cache I'd like to split the amount of units we cache and the amount of units we free, right now they are both CPU_ALLOC_CACHE_SIZE. This means that once we hit that count, we free all of the, and then store the one we were asked to free. That always keeps 1 local, but maybe it'd make more sense to cache just free CPU_ALLOC_CACHE_SIZE/2 (or something like that) so that we retain more than 1 per cpu in case and app preempts when sleeping for IO and the new task on that CPU then issues IO as well. Probably minor. Ran a quick test on nullb0 with 32 sync readers. The test was O_DIRECT on the block device, so I disabled the __blkdev_direct_IO_simple() bypass. With the above branch, we get ~18.0M IOPS, and without we get ~14M IOPS. Both ran with iostats disabled, to avoid any interference from that. -- Jens Axboe -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>
