On Mon, May 18 2009, Hisashi Hifumi wrote: > Hi. > > I wrote a patch that adds blk_run_backing_dev on page_cache_async_readahead > so readahead I/O is unpluged to improve throughput. > > Following is the test result with dd. > > #dd if=testdir/testfile of=/dev/null bs=16384 > > -2.6.30-rc6 > 1048576+0 records in > 1048576+0 records out > 17179869184 bytes (17 GB) copied, 224.182 seconds, 76.6 MB/s > > -2.6.30-rc6-patched > 1048576+0 records in > 1048576+0 records out > 17179869184 bytes (17 GB) copied, 206.465 seconds, 83.2 MB/s > > Sequential read performance on a big file was improved. > Please merge my patch. > > Thanks. > > Signed-off-by: Hisashi Hifumi <hifumi.hisashi@xxxxxxxxxxxxx> > > diff -Nrup linux-2.6.30-rc6.org/mm/readahead.c linux-2.6.30-rc6.unplug/mm/readahead.c > --- linux-2.6.30-rc6.org/mm/readahead.c 2009-05-18 10:46:15.000000000 +0900 > +++ linux-2.6.30-rc6.unplug/mm/readahead.c 2009-05-18 13:00:42.000000000 +0900 > @@ -490,5 +490,7 @@ page_cache_async_readahead(struct addres > > /* do read-ahead */ > ondemand_readahead(mapping, ra, filp, true, offset, req_size); > + > + blk_run_backing_dev(mapping->backing_dev_info, NULL); > } > EXPORT_SYMBOL_GPL(page_cache_async_readahead); I'm surprised this makes much of a difference. It seems correct to me to NOT unplug the device, since it will get unplugged when someone ends up actually waiting for a page. And that will then kick off the remaining IO as well. For this dd case, you'll be hitting lock_page() for the readahead page really soon, definitely not long enough to warrant such a big difference in speed. So, are these numbers 100% reproducible? Could you capture blktrace data for both with and without the patch, so we can take a closer look at the generated IO for each case? -- Jens Axboe -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
