Re: using splice/vmsplice to improve file receive performance

[Jens Axboe <jens.axboe@oracle.com>]

On Wed, Jan 03 2007, saeed bishara wrote:
> On 12/22/06, Jens Axboe <jens.axboe@oracle.com> wrote:
> >On Fri, Dec 22 2006, saeed bishara wrote:
> >> On 12/22/06, Jens Axboe <jens.axboe@oracle.com> wrote:
> >> >On Fri, Dec 22 2006, saeed bishara wrote:
> >> >> On 12/22/06, Jens Axboe <jens.axboe@oracle.com> wrote:
> >> >> >On Thu, Dec 21 2006, saeed bishara wrote:
> >> >> >> Hi,
> >> >> >> I'm trying to use the splice/vmsplice system calls to improve the
> >> >> >> samba server write throughput, but before touching the smbd, I 
> >started
> >> >> >> to improve the ttcp tool since it simple and has the same flow. I'm
> >> >> >> expecting to avoid the "copy_from_user" path when using those
> >> >> >> syscalls.
> >> >> >> so far, I couldn't make any improvement, actually the throughput 
> >get
> >> >> >> worst. the new receive flow looks like this (code also attached):
> >> >> >> 1. read tcp packet (64 pages) to page aligned buffer.
> >> >> >> 2. vmsplice the buffer to pipe with SPLICE_F_MOVE.
> >> >> >> 3. splice the pipe to the file, also with SPLICE_F_MOVE.
> >> >> >>
> >> >> >> the strace shows that the splice takes a lot of time. also when
> >> >> >> profiling the kernel, I found that the memcpy() called to often !!
> >> >> >
> >> >> >(didn't see this until now, axboe@suse.de doesn't work anymore)
> >> >> >
> >> >> >I'm assuming that you mean you vmsplice with SPLICE_F_GIFT, to hand
> >> >> >ownership of the pages to the kernel (in which case SPLICE_F_MOVE 
> >will
> >> >> >work, otherwise you get a copy)? If not, that'll surely cost you a 
> >data
> >> >> >copy
> >> >>   I'll try the vmplice with SPLICE_F_GIFT and splice with MOVE. btw,
> >> >> I noticed that the  splice system call takes the bulk of the time,
> >> >> does it mean anything?
> >> >
> >> >Hard to say without seeing some numbers :-)
> >> I'm out of the office, I'll send it later. btw, my test bed ( the
> >> receiver side ) is arm9. does it matter?
> >
> >The vmsplice is basically vm intensive, so it could matter.
> >
> >> >> >This sounds remarkably like a recent thread on lkml, you may want to
> >> >> >read up on that. Basically using splice for network receive is a bit 
> >of
> >> >> >a work-around now, since you do need the one copy and then vmsplice 
> >that
> >> >> >into a pipe. To realize the full potential of splice, we first need
> >> >> >socket receive support so you can skip that step (splice from socket 
> >to
> >> >> >pipe, splice pipe to file).
> >> >> Ashwini Kulkarni posted patches that implements that, see
> >> >> http://lkml.org/lkml/2006/9/20/272 .  is that right?
> >> >> >
> >> >> >There was no test code attached, btw.
> >> >> sorry, here it is.
> >> >> can you please add sample application to your test tools (splice,fio
> >> >> ,,) that demonstrates my flow; socket to file using read & vmsplice?
> >> >
> >> >I didn't add such an example, since I had hoped that we would have
> >> >splice from socket support sooner rather than later. But I can do so, of
> >> >course.
> >> do you any preliminary patches? I can start playing with it.
> >
> >I don't, Intel posted a set of patches a few months ago though. I didn't
> >have time to look that at the time being, but you should be able to find
> >them in the archives.
> >
> >> >I'll try your test. One thing that sticks out initially is that you
> >> >should be using full pages, the splice pipe will not merge page
> >> >segments. So don't use a buflen less than the page size.
> >>
> >> yes, actually I  run the ttcp with -l65536 ( 64KB ), and the buffer is
> >> always page aligned.also, the splice/vmsplice with MOVE or GIFT will
> >> fail if the buffer is not a whole pages. am I rigth?
> >
> >Yes.
> >
> >I added a simple splice-fromnet example in the splice git repo, see if
> >you can repeat your results with that. Doing:
> >
> ># ./splice-fromnet -g 2001 | ./splice-out -m /dev/null
> >
> >and
> >
> ># cat /dev/zero | netcat localhost 2001
> >
> >gets me about 490MiB/sec, using a recv/write loop is around 413MiB/sec.
> >Not migrating pages gets me around 422MiB/sec.
> >
> >--
> >Jens Axboe
> >
> >
> I've done some investigation in the splice flow and found the following:
> even when using vmsplice with GIFT and splice with MOVE, the user
> buffers still copied, I see that the memcpy from pipe_to_file() is
> called.
> I added debug messages in this function and here what I got:
> 1. the  generic_pipe_buf_steal always fails, this is because the
> page_count is 2.
> 2. after then, the find_lock_page fails as well.
> 3. page_cache_alloc_cold succeeds.
> 4. but, since the buf->page is differs from the page (returned by
> page_cache_alloc_cold) the memcpy function is called.
> 
> this behavior true for all the buffers that vmspliced to ext3 file.
> is this the expected behavior? is there any way to make the steal
> operation return with success?

It works for me, with most pages. Using the vmsplice/splice-out from the
splice tools, doing

$ ./vmsplice -g |  ./splice-out -m g

about half of the pages have count==1 and the steal suceeds.

find_lock_page() will only suceed, if the file exists and is cached
already. splice-out will truncate the file, so it should never suceed
for that case. For both the find_lock_page() success and failure case
(page being allocated), it's a given that we need to copy the data.

-- 
Jens Axboe