On Thu, Feb 12, 2015 at 02:59:45PM -0500, Anna Schumaker wrote: > On 02/11/2015 02:22 PM, Anna Schumaker wrote: > > On 02/11/2015 11:22 AM, J. Bruce Fields wrote: > >> On Wed, Feb 11, 2015 at 11:13:38AM -0500, Trond Myklebust wrote: > >>> On Wed, Feb 11, 2015 at 11:04 AM, Anna Schumaker > >>> <Anna.Schumaker@xxxxxxxxxx> wrote: > >>>> I'm not seeing a huge performance increase with READ_PLUS compared to READ (in fact, it's often a bit slower compared to READ, even when using splice). My guess is that the problem is mostly on the client's end since we have to do a memory shift on each segment to get everything lined up properly. I'm playing around with code that cuts down the number of memory shifts, but I still have a few bugs to work out before I'll know if it actually helps. > >>>> > >>> > >>> I'm wondering if the right way to do READ_PLUS would have been to > >>> instead have a separate function READ_SPARSE, that will return a list > >>> of all sparse areas in the supplied range. We could even make that a > >>> READ_SAME, that can do the same for patterned data. > >> > >> I worry about ending up with incoherent results, but perhaps it's no > >> different from the current behavior since we're already piecing together > >> our idea of the file content from multiple reads sent in parallel. > >> > >>> The thing is that READ works just fine for what we want it to do. The > >>> real win here would be if given a very large file, we could request a > >>> list of all the sparse areas in, say, a 100GB range, and then use that > >>> data to build up a bitmap of unallocated blocks for which we can skip > >>> the READ requests. > >> > >> Can we start by having the server return a single data extent covering > >> the whole read request, with the single exception of the case where the > >> read falls entirely within a hole? > > > > I'm trying this and it's still giving me pretty bad performance. I picked out 6 xfstests that read sparse files, and v4.2 takes almost a minute longer to run compared to v4.1. (1:30 vs 2:22). > > > > I'm going to look into how I zero pages on the client - maybe that can be combined with the right-shift function so pages only need to be mapped into memory once. > > Today I learned all about how to use operf to identify where the bottleneck is :). It looks like the problem is in the hole zeroing code on the client side. Is there a better way than memset() to mark a page as all zeros? I'm a little surprised that replacing a READ with a READ_PLUS + memset is slower, and I'm really surprised that it's a minute slower. Are we sure there isn't something else going on? Does comparing the operation counts (from /proc/self/mountstats) show anything interesting? --b. -- To unsubscribe from this list: send the line "unsubscribe linux-nfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html