On Tue, Jan 07, 2014 at 10:38:46AM -0500, Anna Schumaker wrote: > On 01/07/2014 09:56 AM, J. Bruce Fields wrote: > > On Tue, Jan 07, 2014 at 09:42:04AM -0500, Anna Schumaker wrote: > >> On 01/06/2014 05:32 PM, J. Bruce Fields wrote: > >>> On Mon, Jan 06, 2014 at 04:57:10PM -0500, Anna Schumaker wrote: > >>>> These patches are my initial implementation of READ_PLUS. I still have a > >>>> few issues to work out before they can be applied, but I wanted to submit > >>>> them anyway to get feedback before going much further. These patches were > >>>> developed on top of my earlier SEEK and WRITE_PLUS patches, and probably > >>>> won't apply cleanly without them (I am willing to reorder things if necessary!). > >>>> > >>>> On the server side, I handle the cases where a file is 100% hole, 100% data > >>>> or hole followed by data. Any holes after a data segment will be expanded > >>>> to zeros on the wire. > >>> > >>> I assume that for "a file" I should read "the requested range of the > >>> file"? > >> > >> Yes. > >> > >>> > >>> hole+data+hole should also be doable, shouldn't it? I'd think the real > >>> problem would be multiple data extents. > >> > >> It might be, but I haven't tried it yet. I can soon! > >> > >>> > >>>> This is due to a limitation in the the NFSD > >>>> encode-to-page function that will adjust pointers to point to the xdr tail > >>>> after reading a file to the "pages" section. Bruce, do you have any > >>>> suggestions here? > >>> > >>> The server xdr encoding needs a rewrite. I'll see if I can ignore you > >>> all and put my head down and get a version of that posted this week. > >> > >> :) > >> > >>> > >>> That should make it easier to return all the data, though it will turn > >>> off zero-copy in the case of multiple data extents. > >>> > >>> If we want READ_PLUS to support zero copy in the case of multiple > >>> extents then I think we need a new data structure to represent the > >>> resulting rpc reply. An xdr buf only knows how to insert one array of > >>> pages in the middle of the data. Maybe a list of xdr bufs? > >>> > >>> But that's an annoying job and possibly a premature optimization. > >>> > >>> It might be useful to first understand the typical distribution of holes > >>> in a file and how likely various workloads are to produce reads with > >>> multiple holes in the middle. > >> > >> I already have a few performance numbers, but nothing that can be trusted due to the number of debugging printk()s I used to make sure the client decoded everything correctly. My plan is to collect the following information using: v4.0, v4.1, v4.2 (SEEK), v4.2 (SEEK + WRITE_PLUS), and v4.2 (SEEK + WRITE_PLUS + READ_PLUS). > > > > What's the workload and hardware setup? > > I was going to run filebench tests (fileserver, mongo, varmail) between two VMs. I only have the one laptop with me today, so I can't test between two real machines without asking for a volunteer from Workantile. I am planning to kill Firefox and Thunderbird before running anything! That doesn't seem like an interesting test. Running some generic filesystem benchmarks sounds like a fine idea. If nothing else to check for regressions. But if we want to figure out whether it helps where it's supposed to then I think we need to think about what it's meant to do (reduce bandwidth use transferring sparse files, I guess. Maybe something copying VM images would be 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
