[00/16] libceph: messenger: send/recv data at one go
mbox series

Message ID 20200421131850.443228-1-rpenyaev@suse.de
Headers show
Series
  • libceph: messenger: send/recv data at one go
Related show

Message

Roman Penyaev April 21, 2020, 1:18 p.m. UTC
Hi folks,

While experimenting with messenger code in userspace [1] I noticed
that send and receive socket calls always operate with 4k, even bvec
length is larger (for example when bvec is contructed from bio, where
multi-page is used for big IOs). This is an attempt to speed up send
and receive for large IO.

First 3 patches are cleanups. I remove unused code and get rid of
ceph_osd_data structure. I found that ceph_osd_data duplicates
ceph_msg_data and it seems unified API looks better for similar
things.

In the following patches ceph_msg_data_cursor is switched to iov_iter,
which seems is more suitable for such kind of things (when we
basically do socket IO). This gives us the possibility to use the
whole iov_iter for sendmsg() and recvmsg() calls instead of iterating
page by page. sendpage() call also benefits from this, because now if
bvec is constructed from multi-page, then we can 0-copy the whole
bvec in one go.

I also allowed myself to get rid of ->last_piece and ->need_crc
members and ceph_msg_data_next() call. Now CRC is calculated not on
page basis, but according to the size of processed chunk.  I found
ceph_msg_data_next() is a bit redundant, since we always can set the
next cursor chunk on cursor init or on advance.

How I tested the performance? I used rbd.fio load on 1 OSD in memory
with the following fio configuration:

  direct=1
  time_based=1
  runtime=10
  ioengine=io_uring
  size=256m

  rw=rand{read|write}
  numjobs=32
  iodepth=32

  [job1]
  filename=/dev/rbd0

RBD device is mapped with 'nocrc' option set.  For writes OSD completes
requests immediately, without touching the memory simulating null block
device, that's why write throughput in my results is much higher than
for reads.

I tested on loopback interface only, in Vm, have not yet setup the
cluster on real machines, so sendpage() on a big multi-page shows
indeed good results, as expected. But I found an interesting comment
in drivers/infiniband/sw/siw/siw_qp_tc.c:siw_tcp_sendpages(), which
says:

 "Using sendpage to push page by page appears to be less efficient
  than using sendmsg, even if data are copied.
 
  A general performance limitation might be the extra four bytes
  trailer checksum segment to be pushed after user data."

I could not prove or disprove since have tested on loopback interface
only.  So it might be that sendmsg() in on go is faster than
sendpage() for bvecs with many segments.

Here is the output of the rbd fio load for various block sizes:

==== WRITE ===

current master, rw=randwrite, numjobs=32 iodepth=32

  4k  IOPS=92.7k, BW=362MiB/s, Lat=11033.30usec
  8k  IOPS=85.6k, BW=669MiB/s, Lat=11956.74usec
 16k  IOPS=76.8k, BW=1200MiB/s, Lat=13318.24usec
 32k  IOPS=56.7k, BW=1770MiB/s, Lat=18056.92usec
 64k  IOPS=34.0k, BW=2186MiB/s, Lat=29.23msec
128k  IOPS=21.8k, BW=2720MiB/s, Lat=46.96msec
256k  IOPS=14.4k, BW=3596MiB/s, Lat=71.03msec
512k  IOPS=8726, BW=4363MiB/s, Lat=116.34msec
  1m  IOPS=4799, BW=4799MiB/s, Lat=211.15msec

this patchset,  rw=randwrite, numjobs=32 iodepth=32

  4k  IOPS=94.7k, BW=370MiB/s, Lat=10802.43usec
  8k  IOPS=91.2k, BW=712MiB/s, Lat=11221.00usec
 16k  IOPS=80.4k, BW=1257MiB/s, Lat=12715.56usec
 32k  IOPS=61.2k, BW=1912MiB/s, Lat=16721.33usec
 64k  IOPS=40.9k, BW=2554MiB/s, Lat=24993.31usec
128k  IOPS=25.7k, BW=3216MiB/s, Lat=39.72msec
256k  IOPS=17.3k, BW=4318MiB/s, Lat=59.15msec
512k  IOPS=11.1k, BW=5559MiB/s, Lat=91.39msec
  1m  IOPS=6696, BW=6696MiB/s, Lat=151.25msec


=== READ ===

current master, rw=randread, numjobs=32 iodepth=32

  4k  IOPS=62.5k, BW=244MiB/s, Lat=16.38msec
  8k  IOPS=55.5k, BW=433MiB/s, Lat=18.44msec
 16k  IOPS=40.6k, BW=635MiB/s, Lat=25.18msec
 32k  IOPS=24.6k, BW=768MiB/s, Lat=41.61msec
 64k  IOPS=14.8k, BW=925MiB/s, Lat=69.06msec
128k  IOPS=8687, BW=1086MiB/s, Lat=117.59msec
256k  IOPS=4733, BW=1183MiB/s, Lat=214.76msec
512k  IOPS=3156, BW=1578MiB/s, Lat=320.54msec
  1m  IOPS=1901, BW=1901MiB/s, Lat=528.22msec

this patchset,  rw=randread, numjobs=32 iodepth=32

  4k  IOPS=62.6k, BW=244MiB/s, Lat=16342.89usec
  8k  IOPS=55.5k, BW=434MiB/s, Lat=18.42msec
 16k  IOPS=43.2k, BW=675MiB/s, Lat=23.68msec
 32k  IOPS=28.4k, BW=887MiB/s, Lat=36.04msec
 64k  IOPS=20.2k, BW=1263MiB/s, Lat=50.54msec
128k  IOPS=11.7k, BW=1465MiB/s, Lat=87.01msec
256k  IOPS=6813, BW=1703MiB/s, Lat=149.30msec
512k  IOPS=5363, BW=2682MiB/s, Lat=189.37msec
  1m  IOPS=2220, BW=2221MiB/s, Lat=453.92msec


Results for small blocks are not interesting, since there should not
be any difference. But starting from 32k block benefits of doing IO
for the whole message at once starts to prevail.

I'm open to test any other loads, I just usually stick to fio rbd,
since it is pretty simple and pumps the IOs quite well.

[1] https://github.com/rouming/pech

Roman Penyaev (16):
  libceph: remove unused ceph_pagelist_cursor
  libceph: extend ceph_msg_data API in order to switch on it
  libceph,rbd,cephfs: switch from ceph_osd_data to ceph_msg_data
  libceph: remove ceph_osd_data completely
  libceph: remove unused last_piece out parameter from
    ceph_msg_data_next()
  libceph: switch data cursor from page to iov_iter for messenger
  libceph: use new tcp_sendiov() instead of tcp_sendmsg() for messenger
  libceph: remove unused tcp wrappers, now iov_iter is used for
    messenger
  libceph: no need for cursor->need_crc for messenger
  libceph: remove ->last_piece member for message data cursor
  libceph: remove not necessary checks on doing advance on bio and bvecs
    cursor
  libceph: switch bvecs cursor to iov_iter for messenger
  libceph: switch bio cursor to iov_iter for messenger
  libceph: switch pages cursor to iov_iter for messenger
  libceph: switch pageslist cursor to iov_iter for messenger
  libceph: remove ceph_msg_data_*_next() from messenger

 drivers/block/rbd.c             |   4 +-
 fs/ceph/addr.c                  |  10 +-
 fs/ceph/file.c                  |   4 +-
 include/linux/ceph/messenger.h  |  42 ++-
 include/linux/ceph/osd_client.h |  58 +---
 include/linux/ceph/pagelist.h   |  12 -
 net/ceph/messenger.c            | 558 +++++++++++++++-----------------
 net/ceph/osd_client.c           | 251 ++++----------
 net/ceph/pagelist.c             |  38 ---
 9 files changed, 390 insertions(+), 587 deletions(-)

Comments

Ilya Dryomov April 21, 2020, 3:51 p.m. UTC | #1
On Tue, Apr 21, 2020 at 3:18 PM Roman Penyaev <rpenyaev@suse.de> wrote:
>
> Hi folks,
>
> While experimenting with messenger code in userspace [1] I noticed
> that send and receive socket calls always operate with 4k, even bvec
> length is larger (for example when bvec is contructed from bio, where
> multi-page is used for big IOs). This is an attempt to speed up send
> and receive for large IO.
>
> First 3 patches are cleanups. I remove unused code and get rid of
> ceph_osd_data structure. I found that ceph_osd_data duplicates
> ceph_msg_data and it seems unified API looks better for similar
> things.
>
> In the following patches ceph_msg_data_cursor is switched to iov_iter,
> which seems is more suitable for such kind of things (when we
> basically do socket IO). This gives us the possibility to use the
> whole iov_iter for sendmsg() and recvmsg() calls instead of iterating
> page by page. sendpage() call also benefits from this, because now if
> bvec is constructed from multi-page, then we can 0-copy the whole
> bvec in one go.

Hi Roman,

I'm in the process of rewriting the kernel messenger to support msgr2
(i.e. encryption) and noticed the same things.  The switch to iov_iter
was the first thing I implemented ;)  Among other things is support for
multipage bvecs and explicit socket corking.  I haven't benchmarked any
of it though -- it just seemed like a sensible thing to do, especially
since the sendmsg/sendpage infrastructure needed changes for encryption
anyway.

Support for kvecs isn't implemented yet, but will be in order to get
rid of all those "allocate a page just to process 16 bytes" sites.

Unfortunately I got distracted by some higher priority issues with the
userspace messenger, so the kernel messenger is in a bit of a state of
disarray at the moment.  Here is the excerpt from the send path:

#define CEPH_MSG_FLAGS (MSG_DONTWAIT | MSG_NOSIGNAL)

static int do_sendmsg(struct ceph_connection *con, struct iov_iter *it)
{
        struct msghdr msg = { .msg_flags = CEPH_MSG_FLAGS };
        int ret;

        msg.msg_iter = *it;
        while (iov_iter_count(it)) {
                ret = do_one_sendmsg(con, &msg);
                if (ret <= 0) {
                        if (ret == -EAGAIN)
                                ret = 0;
                        return ret;
                }

                iov_iter_advance(it, ret);
        }

        BUG_ON(msg_data_left(&msg));
        return 1;
}

static int do_sendpage(struct ceph_connection *con, struct iov_iter *it)
{
        ssize_t ret;

        BUG_ON(!iov_iter_is_bvec(it));
        while (iov_iter_count(it)) {
                struct page *page = it->bvec->bv_page;
                int offset = it->bvec->bv_offset + it->iov_offset;
                size_t size = min(it->count,
                                  it->bvec->bv_len - it->iov_offset);

                /*
                 * sendpage cannot properly handle pages with
                 * page_count == 0, we need to fall back to sendmsg if
                 * that's the case.
                 *
                 * Same goes for slab pages: skb_can_coalesce() allows
                 * coalescing neighboring slab objects into a single frag
                 * which triggers one of hardened usercopy checks.
                 */
                if (page_count(page) >= 1 && !PageSlab(page)) {
                        ret = do_one_sendpage(con, page, offset, size,
                                              CEPH_MSG_FLAGS);
                } else {
                        struct msghdr msg = { .msg_flags = CEPH_MSG_FLAGS };
                        struct bio_vec bv = {
                                .bv_page = page,
                                .bv_offset = offset,
                                .bv_len = size,
                        };

                        iov_iter_bvec(&msg.msg_iter, WRITE, &bv, 1, size);
                        ret = do_one_sendmsg(con, &msg);
                }
                if (ret <= 0) {
                        if (ret == -EAGAIN)
                                ret = 0;
                        return ret;
                }

                iov_iter_advance(it, ret);
        }

        return 1;
}

/*
 * Write as much as possible.  The socket is expected to be corked,
 * so we don't bother with MSG_MORE/MSG_SENDPAGE_NOTLAST here.
 *
 * Return:
 *   1 - done, nothing else to write
 *   0 - socket is full, need to wait
 *  <0 - error
 */
int ceph_tcp_send(struct ceph_connection *con)
{
        bool is_kvec = iov_iter_is_kvec(&con->out_iter);
        int ret;

        dout("%s con %p have %zu is_kvec %d\n", __func__, con,
             iov_iter_count(&con->out_iter), is_kvec);
        if (is_kvec)
                ret = do_sendmsg(con, &con->out_iter);
        else
                ret = do_sendpage(con, &con->out_iter);

        dout("%s con %p ret %d left %zu\n", __func__, con, ret,
             iov_iter_count(&con->out_iter));
        return ret;
}

I'll make sure to CC you on my patches, should be in a few weeks.

Getting rid of ceph_osd_data is probably a good idea.  FWIW I never
liked it, but not strong enough to bother with removing it.

>
> I also allowed myself to get rid of ->last_piece and ->need_crc
> members and ceph_msg_data_next() call. Now CRC is calculated not on
> page basis, but according to the size of processed chunk.  I found
> ceph_msg_data_next() is a bit redundant, since we always can set the
> next cursor chunk on cursor init or on advance.
>
> How I tested the performance? I used rbd.fio load on 1 OSD in memory
> with the following fio configuration:
>
>   direct=1
>   time_based=1
>   runtime=10
>   ioengine=io_uring
>   size=256m
>
>   rw=rand{read|write}
>   numjobs=32
>   iodepth=32
>
>   [job1]
>   filename=/dev/rbd0
>
> RBD device is mapped with 'nocrc' option set.  For writes OSD completes
> requests immediately, without touching the memory simulating null block
> device, that's why write throughput in my results is much higher than
> for reads.
>
> I tested on loopback interface only, in Vm, have not yet setup the
> cluster on real machines, so sendpage() on a big multi-page shows
> indeed good results, as expected. But I found an interesting comment
> in drivers/infiniband/sw/siw/siw_qp_tc.c:siw_tcp_sendpages(), which
> says:
>
>  "Using sendpage to push page by page appears to be less efficient
>   than using sendmsg, even if data are copied.
>
>   A general performance limitation might be the extra four bytes
>   trailer checksum segment to be pushed after user data."
>
> I could not prove or disprove since have tested on loopback interface
> only.  So it might be that sendmsg() in on go is faster than
> sendpage() for bvecs with many segments.

Please share any further findings.  We have been using sendpage for
the data section of the message since forever and I remember hearing
about a performance regression when someone inadvertently disabled the
sendpage path (can't recall the subsystem -- iSCSI?).  If you discover
that sendpage is actually slower, that would be very interesting.

>
> Here is the output of the rbd fio load for various block sizes:
>
> ==== WRITE ===
>
> current master, rw=randwrite, numjobs=32 iodepth=32
>
>   4k  IOPS=92.7k, BW=362MiB/s, Lat=11033.30usec
>   8k  IOPS=85.6k, BW=669MiB/s, Lat=11956.74usec
>  16k  IOPS=76.8k, BW=1200MiB/s, Lat=13318.24usec
>  32k  IOPS=56.7k, BW=1770MiB/s, Lat=18056.92usec
>  64k  IOPS=34.0k, BW=2186MiB/s, Lat=29.23msec
> 128k  IOPS=21.8k, BW=2720MiB/s, Lat=46.96msec
> 256k  IOPS=14.4k, BW=3596MiB/s, Lat=71.03msec
> 512k  IOPS=8726, BW=4363MiB/s, Lat=116.34msec
>   1m  IOPS=4799, BW=4799MiB/s, Lat=211.15msec
>
> this patchset,  rw=randwrite, numjobs=32 iodepth=32
>
>   4k  IOPS=94.7k, BW=370MiB/s, Lat=10802.43usec
>   8k  IOPS=91.2k, BW=712MiB/s, Lat=11221.00usec
>  16k  IOPS=80.4k, BW=1257MiB/s, Lat=12715.56usec
>  32k  IOPS=61.2k, BW=1912MiB/s, Lat=16721.33usec
>  64k  IOPS=40.9k, BW=2554MiB/s, Lat=24993.31usec
> 128k  IOPS=25.7k, BW=3216MiB/s, Lat=39.72msec
> 256k  IOPS=17.3k, BW=4318MiB/s, Lat=59.15msec
> 512k  IOPS=11.1k, BW=5559MiB/s, Lat=91.39msec
>   1m  IOPS=6696, BW=6696MiB/s, Lat=151.25msec
>
>
> === READ ===
>
> current master, rw=randread, numjobs=32 iodepth=32
>
>   4k  IOPS=62.5k, BW=244MiB/s, Lat=16.38msec
>   8k  IOPS=55.5k, BW=433MiB/s, Lat=18.44msec
>  16k  IOPS=40.6k, BW=635MiB/s, Lat=25.18msec
>  32k  IOPS=24.6k, BW=768MiB/s, Lat=41.61msec
>  64k  IOPS=14.8k, BW=925MiB/s, Lat=69.06msec
> 128k  IOPS=8687, BW=1086MiB/s, Lat=117.59msec
> 256k  IOPS=4733, BW=1183MiB/s, Lat=214.76msec
> 512k  IOPS=3156, BW=1578MiB/s, Lat=320.54msec
>   1m  IOPS=1901, BW=1901MiB/s, Lat=528.22msec
>
> this patchset,  rw=randread, numjobs=32 iodepth=32
>
>   4k  IOPS=62.6k, BW=244MiB/s, Lat=16342.89usec
>   8k  IOPS=55.5k, BW=434MiB/s, Lat=18.42msec
>  16k  IOPS=43.2k, BW=675MiB/s, Lat=23.68msec
>  32k  IOPS=28.4k, BW=887MiB/s, Lat=36.04msec
>  64k  IOPS=20.2k, BW=1263MiB/s, Lat=50.54msec
> 128k  IOPS=11.7k, BW=1465MiB/s, Lat=87.01msec
> 256k  IOPS=6813, BW=1703MiB/s, Lat=149.30msec
> 512k  IOPS=5363, BW=2682MiB/s, Lat=189.37msec
>   1m  IOPS=2220, BW=2221MiB/s, Lat=453.92msec
>
>
> Results for small blocks are not interesting, since there should not
> be any difference. But starting from 32k block benefits of doing IO
> for the whole message at once starts to prevail.

It's not really the whole message, just the header, front and middle
sections, right?  The data section is still per-bvec, it's just that
bvec is no longer limited to a single page but may encompass several
physically contiguous pages.  These are not that easy to come by on
a heavily loaded system, but they do result in nice numbers.

Thanks,

                Ilya
Roman Penyaev April 21, 2020, 4:28 p.m. UTC | #2
Hi Ilya,

On 2020-04-21 17:51, Ilya Dryomov wrote:
> On Tue, Apr 21, 2020 at 3:18 PM Roman Penyaev <rpenyaev@suse.de> wrote:
>> 
>> Hi folks,
>> 
>> While experimenting with messenger code in userspace [1] I noticed
>> that send and receive socket calls always operate with 4k, even bvec
>> length is larger (for example when bvec is contructed from bio, where
>> multi-page is used for big IOs). This is an attempt to speed up send
>> and receive for large IO.
>> 
>> First 3 patches are cleanups. I remove unused code and get rid of
>> ceph_osd_data structure. I found that ceph_osd_data duplicates
>> ceph_msg_data and it seems unified API looks better for similar
>> things.
>> 
>> In the following patches ceph_msg_data_cursor is switched to iov_iter,
>> which seems is more suitable for such kind of things (when we
>> basically do socket IO). This gives us the possibility to use the
>> whole iov_iter for sendmsg() and recvmsg() calls instead of iterating
>> page by page. sendpage() call also benefits from this, because now if
>> bvec is constructed from multi-page, then we can 0-copy the whole
>> bvec in one go.
> 
> Hi Roman,
> 
> I'm in the process of rewriting the kernel messenger to support msgr2
> (i.e. encryption) and noticed the same things.  The switch to iov_iter
> was the first thing I implemented ;)  Among other things is support for
> multipage bvecs and explicit socket corking.

Ah, ok, good to know. This patchset came from the userspace variant of
the kernel messenger. These changes also show nice numbers on userspace
side. (of course without sendpage() variant and plus some caching on
receive, which I also implemented for kernel side, but left aside since
this did not show any interesting results for rbd load)

> I haven't benchmarked any
> of it though -- it just seemed like a sensible thing to do, especially
> since the sendmsg/sendpage infrastructure needed changes for encryption
> anyway.

I can benchmark on my localhost setup easily. Just add me to CC when
you are done.

> 
> Support for kvecs isn't implemented yet, but will be in order to get
> rid of all those "allocate a page just to process 16 bytes" sites.
> 
> Unfortunately I got distracted by some higher priority issues with the
> userspace messenger, so the kernel messenger is in a bit of a state of
> disarray at the moment.  Here is the excerpt from the send path:
> 
> #define CEPH_MSG_FLAGS (MSG_DONTWAIT | MSG_NOSIGNAL)
> 
> static int do_sendmsg(struct ceph_connection *con, struct iov_iter *it)
> {
>         struct msghdr msg = { .msg_flags = CEPH_MSG_FLAGS };
>         int ret;
> 
>         msg.msg_iter = *it;
>         while (iov_iter_count(it)) {
>                 ret = do_one_sendmsg(con, &msg);
>                 if (ret <= 0) {
>                         if (ret == -EAGAIN)
>                                 ret = 0;
>                         return ret;
>                 }
> 
>                 iov_iter_advance(it, ret);
>         }
> 
>         BUG_ON(msg_data_left(&msg));
>         return 1;
> }
> 
> static int do_sendpage(struct ceph_connection *con, struct iov_iter 
> *it)
> {
>         ssize_t ret;
> 
>         BUG_ON(!iov_iter_is_bvec(it));
>         while (iov_iter_count(it)) {
>                 struct page *page = it->bvec->bv_page;
>                 int offset = it->bvec->bv_offset + it->iov_offset;
>                 size_t size = min(it->count,
>                                   it->bvec->bv_len - it->iov_offset);
> 
>                 /*
>                  * sendpage cannot properly handle pages with
>                  * page_count == 0, we need to fall back to sendmsg if
>                  * that's the case.
>                  *
>                  * Same goes for slab pages: skb_can_coalesce() allows
>                  * coalescing neighboring slab objects into a single 
> frag
>                  * which triggers one of hardened usercopy checks.
>                  */
>                 if (page_count(page) >= 1 && !PageSlab(page)) {
>                         ret = do_one_sendpage(con, page, offset, size,
>                                               CEPH_MSG_FLAGS);
>                 } else {
>                         struct msghdr msg = { .msg_flags = 
> CEPH_MSG_FLAGS };
>                         struct bio_vec bv = {
>                                 .bv_page = page,
>                                 .bv_offset = offset,
>                                 .bv_len = size,
>                         };
> 
>                         iov_iter_bvec(&msg.msg_iter, WRITE, &bv, 1, 
> size);
>                         ret = do_one_sendmsg(con, &msg);
>                 }
>                 if (ret <= 0) {
>                         if (ret == -EAGAIN)
>                                 ret = 0;
>                         return ret;
>                 }
> 
>                 iov_iter_advance(it, ret);
>         }
> 
>         return 1;
> }
> 
> /*
>  * Write as much as possible.  The socket is expected to be corked,
>  * so we don't bother with MSG_MORE/MSG_SENDPAGE_NOTLAST here.
>  *
>  * Return:
>  *   1 - done, nothing else to write
>  *   0 - socket is full, need to wait
>  *  <0 - error
>  */
> int ceph_tcp_send(struct ceph_connection *con)
> {
>         bool is_kvec = iov_iter_is_kvec(&con->out_iter);
>         int ret;
> 
>         dout("%s con %p have %zu is_kvec %d\n", __func__, con,
>              iov_iter_count(&con->out_iter), is_kvec);
>         if (is_kvec)
>                 ret = do_sendmsg(con, &con->out_iter);
>         else
>                 ret = do_sendpage(con, &con->out_iter);
> 
>         dout("%s con %p ret %d left %zu\n", __func__, con, ret,
>              iov_iter_count(&con->out_iter));
>         return ret;
> }

Ha! Nice! That almost exactly what I do in current patchset, except
corking. I still bother with MSG_MORE/MSG_SENDPAGE_NOTLAST :)

BTW kvecs also can be sendpaged. Since we do not have userspace
iovec, thus page can be easily taken.  But that is a minor.

> I'll make sure to CC you on my patches, should be in a few weeks.

Yes, please.  I can help with benchmarking and reviewing.

> 
> Getting rid of ceph_osd_data is probably a good idea.  FWIW I never
> liked it, but not strong enough to bother with removing it.
> 
>> 
>> I also allowed myself to get rid of ->last_piece and ->need_crc
>> members and ceph_msg_data_next() call. Now CRC is calculated not on
>> page basis, but according to the size of processed chunk.  I found
>> ceph_msg_data_next() is a bit redundant, since we always can set the
>> next cursor chunk on cursor init or on advance.
>> 
>> How I tested the performance? I used rbd.fio load on 1 OSD in memory
>> with the following fio configuration:
>> 
>>   direct=1
>>   time_based=1
>>   runtime=10
>>   ioengine=io_uring
>>   size=256m
>> 
>>   rw=rand{read|write}
>>   numjobs=32
>>   iodepth=32
>> 
>>   [job1]
>>   filename=/dev/rbd0
>> 
>> RBD device is mapped with 'nocrc' option set.  For writes OSD 
>> completes
>> requests immediately, without touching the memory simulating null 
>> block
>> device, that's why write throughput in my results is much higher than
>> for reads.
>> 
>> I tested on loopback interface only, in Vm, have not yet setup the
>> cluster on real machines, so sendpage() on a big multi-page shows
>> indeed good results, as expected. But I found an interesting comment
>> in drivers/infiniband/sw/siw/siw_qp_tc.c:siw_tcp_sendpages(), which
>> says:
>> 
>>  "Using sendpage to push page by page appears to be less efficient
>>   than using sendmsg, even if data are copied.
>> 
>>   A general performance limitation might be the extra four bytes
>>   trailer checksum segment to be pushed after user data."
>> 
>> I could not prove or disprove since have tested on loopback interface
>> only.  So it might be that sendmsg() in on go is faster than
>> sendpage() for bvecs with many segments.
> 
> Please share any further findings.  We have been using sendpage for
> the data section of the message since forever and I remember hearing
> about a performance regression when someone inadvertently disabled the
> sendpage path (can't recall the subsystem -- iSCSI?).  If you discover
> that sendpage is actually slower, that would be very interesting.

I will try to find a good machine for that with fat network.
Should be easy to benchmark.

>> Here is the output of the rbd fio load for various block sizes:
>> 
>> ==== WRITE ===
>> 
>> current master, rw=randwrite, numjobs=32 iodepth=32
>> 
>>   4k  IOPS=92.7k, BW=362MiB/s, Lat=11033.30usec
>>   8k  IOPS=85.6k, BW=669MiB/s, Lat=11956.74usec
>>  16k  IOPS=76.8k, BW=1200MiB/s, Lat=13318.24usec
>>  32k  IOPS=56.7k, BW=1770MiB/s, Lat=18056.92usec
>>  64k  IOPS=34.0k, BW=2186MiB/s, Lat=29.23msec
>> 128k  IOPS=21.8k, BW=2720MiB/s, Lat=46.96msec
>> 256k  IOPS=14.4k, BW=3596MiB/s, Lat=71.03msec
>> 512k  IOPS=8726, BW=4363MiB/s, Lat=116.34msec
>>   1m  IOPS=4799, BW=4799MiB/s, Lat=211.15msec
>> 
>> this patchset,  rw=randwrite, numjobs=32 iodepth=32
>> 
>>   4k  IOPS=94.7k, BW=370MiB/s, Lat=10802.43usec
>>   8k  IOPS=91.2k, BW=712MiB/s, Lat=11221.00usec
>>  16k  IOPS=80.4k, BW=1257MiB/s, Lat=12715.56usec
>>  32k  IOPS=61.2k, BW=1912MiB/s, Lat=16721.33usec
>>  64k  IOPS=40.9k, BW=2554MiB/s, Lat=24993.31usec
>> 128k  IOPS=25.7k, BW=3216MiB/s, Lat=39.72msec
>> 256k  IOPS=17.3k, BW=4318MiB/s, Lat=59.15msec
>> 512k  IOPS=11.1k, BW=5559MiB/s, Lat=91.39msec
>>   1m  IOPS=6696, BW=6696MiB/s, Lat=151.25msec
>> 
>> 
>> === READ ===
>> 
>> current master, rw=randread, numjobs=32 iodepth=32
>> 
>>   4k  IOPS=62.5k, BW=244MiB/s, Lat=16.38msec
>>   8k  IOPS=55.5k, BW=433MiB/s, Lat=18.44msec
>>  16k  IOPS=40.6k, BW=635MiB/s, Lat=25.18msec
>>  32k  IOPS=24.6k, BW=768MiB/s, Lat=41.61msec
>>  64k  IOPS=14.8k, BW=925MiB/s, Lat=69.06msec
>> 128k  IOPS=8687, BW=1086MiB/s, Lat=117.59msec
>> 256k  IOPS=4733, BW=1183MiB/s, Lat=214.76msec
>> 512k  IOPS=3156, BW=1578MiB/s, Lat=320.54msec
>>   1m  IOPS=1901, BW=1901MiB/s, Lat=528.22msec
>> 
>> this patchset,  rw=randread, numjobs=32 iodepth=32
>> 
>>   4k  IOPS=62.6k, BW=244MiB/s, Lat=16342.89usec
>>   8k  IOPS=55.5k, BW=434MiB/s, Lat=18.42msec
>>  16k  IOPS=43.2k, BW=675MiB/s, Lat=23.68msec
>>  32k  IOPS=28.4k, BW=887MiB/s, Lat=36.04msec
>>  64k  IOPS=20.2k, BW=1263MiB/s, Lat=50.54msec
>> 128k  IOPS=11.7k, BW=1465MiB/s, Lat=87.01msec
>> 256k  IOPS=6813, BW=1703MiB/s, Lat=149.30msec
>> 512k  IOPS=5363, BW=2682MiB/s, Lat=189.37msec
>>   1m  IOPS=2220, BW=2221MiB/s, Lat=453.92msec
>> 
>> 
>> Results for small blocks are not interesting, since there should not
>> be any difference. But starting from 32k block benefits of doing IO
>> for the whole message at once starts to prevail.
> 
> It's not really the whole message, just the header, front and middle
> sections, right?

No, that is the whole message. Output of fio rbd load. Or I did not get
your question.

> The data section is still per-bvec, it's just that
> bvec is no longer limited to a single page but may encompass several
> physically contiguous pages.

True. Data section is bvec, which is taken from a bio, which on
its turn has one big physically contiguous multipage.  Of course
when there is such a slice of physical contiguous memory.

> These are not that easy to come by on
> a heavily loaded system, but they do result in nice numbers.

Yeah, sometimes there is quite a big number of segments for a big IO.
And for that case it seems makes sense to do buffering, i.e. calling
sendmsg().

So probably sendpage() makes sense to call when

    nr_segs < N && bvec->bv_len > M * 4k

where N and M are some magic numbers which help to resuce costs
of calling do_tcp_sendpages() in a loooong loop. But this is
pure speculation.

--
Roman