Re: 2.6.30-rc deadline scheduler performance regression for iozone over NFS

[Olga Kornievskaia <aglo@citi.umich.edu>]

I believe what you are seeing is how well TCP autotuning performs.
What old NFS code was doing is disabling autotuning and instead using
#nfsd thread to scale TCP recv window. You are providing an example of
where setting TCP buffer sizes outperforms TCP autotuning. While this
is a valid example, there is also an alternative example of where old
NFS design hurts performance.

On Wed, May 13, 2009 at 10:58 AM, Jeff Moyer <jmoyer@redhat.com> wrote:
> Andrew Morton <akpm@linux-foundation.org> writes:
>
>> (obvious cc's added...)
>>
>> It's an iozone performance regression.
>>
>> On Tue, 12 May 2009 23:29:30 -0400 Jeff Moyer <jmoyer@redhat.com> wrote:
>>
>>> Jens Axboe <jens.axboe@oracle.com> writes:
>>>
>>> > On Mon, May 11 2009, Jeff Moyer wrote:
>>> >> Jens Axboe <jens.axboe@oracle.com> writes:
>>> >>
>>> >> > On Fri, May 08 2009, Andrew Morton wrote:
>>> >> >> On Thu, 23 Apr 2009 10:01:58 -0400
>>> >> >> Jeff Moyer <jmoyer@redhat.com> wrote:
>>> >> >>
>>> >> >> > Hi,
>>> >> >> >
>>> >> >> > I've been working on CFQ improvements for interleaved I/Os between
>>> >> >> > processes, and noticed a regression in performance when using the
>>> >> >> > deadline I/O scheduler.  The test uses a server configured with a cciss
>>> >> >> > array and 1Gb/s ethernet.
>>> >> >> >
>>> >> >> > The iozone command line was:
>>> >> >> >   iozone -s 2000000 -r 64 -f /mnt/test/testfile -i 1 -w
>>> >> >> >
>>> >> >> > The numbers in the nfsd's row represent the number of nfsd "threads".
>>> >> >> > These numbers (in MB/s) represent the average of 5 runs.
>>> >> >> >
>>> >> >> >                v2.6.29
>>> >> >> >
>>> >> >> > nfsd's  |   1    |  2   |   4   |   8
>>> >> >> > --------+---------------+-------+------
>>> >> >> > deadline| 43207 | 67436 | 96289 | 107590
>>> >> >> >
>>> >> >> >               2.6.30-rc1
>>> >> >> >
>>> >> >> > nfsd's  |   1   |   2   |   4   |   8
>>> >> >> > --------+---------------+-------+------
>>> >> >> > deadline| 43732 | 68059 | 76659 | 83231
>>> >> >> >
>>> >> >> >     2.6.30-rc3.block-for-linus
>>> >> >> >
>>> >> >> > nfsd's  |   1   |   2   |   4   |   8
>>> >> >> > --------+---------------+-------+------
>>> >> >> > deadline| 46102 | 71151 | 83120 | 82330
>>> >> >> >
>>> >> >> >
>>> >> >> > Notice the drop for 4 and 8 threads.  It may be worth noting that the
>>> >> >> > default number of NFSD threads is 8.
>
> Just following up with numbers:
>
>  2.6.30-rc4
>
> nfsd's  |   8
> --------+------
> cfq     | 51632   (49791 52436 52308 51488 52141)
> deadline| 65558   (41675 42559 74820 87518 81221)
>
>   2.6.30-rc4 reverting the sunrpc "fix"
>
> nfsd's  |   8
> --------+------
> cfq     |  82513  (81650 82762 83147 82935 82073)
> deadline| 107827  (109730 106077 107175 108524 107632)
>
> The numbers in parenthesis are the individual runs.  Notice how
> 2.6.30-rc4 has some pretty wide variations for deadline.
>
> Cheers,
> Jeff
>
>>> >> >> I guess we should ask Rafael to add this to the post-2.6.29 regression
>>> >> >> list.
>>> >> >
>>> >> > I agree. It'd be nice to bisect this one down, I'm guessing some mm
>>> >> > change has caused this writeout regression.
>>> >>
>>> >> It's not writeout, it's a read test.
>>> >
>>> > Doh sorry, I even ran these tests as well a few weeks back. So perhaps
>>> > some read-ahead change, I didn't look into it. FWIW, on a single SATA
>>> > drive here, it didn't show any difference.
>>>
>>> OK, I bisected this to the following commit.  The mount is done using
>>> NFSv3, by the way.
>>>
>>> commit 47a14ef1af48c696b214ac168f056ddc79793d0e
>>> Author: Olga Kornievskaia <aglo@citi.umich.edu>
>>> Date:   Tue Oct 21 14:13:47 2008 -0400
>>>
>>>     svcrpc: take advantage of tcp autotuning
>>>
>>>     Allow the NFSv4 server to make use of TCP autotuning behaviour, which
>>>     was previously disabled by setting the sk_userlocks variable.
>>>
>>>     Set the receive buffers to be big enough to receive the whole RPC
>>>     request, and set this for the listening socket, not the accept socket.
>>>
>>>     Remove the code that readjusts the receive/send buffer sizes for the
>>>     accepted socket. Previously this code was used to influence the TCP
>>>     window management behaviour, which is no longer needed when autotuning
>>>     is enabled.
>>>
>>>     This can improve IO bandwidth on networks with high bandwidth-delay
>>>     products, where a large tcp window is required.  It also simplifies
>>>     performance tuning, since getting adequate tcp buffers previously
>>>     required increasing the number of nfsd threads.
>>>
>>>     Signed-off-by: Olga Kornievskaia <aglo@citi.umich.edu>
>>>     Cc: Jim Rees <rees@umich.edu>
>>>     Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
>>>
>>> diff --git a/net/sunrpc/svcsock.c b/net/sunrpc/svcsock.c
>>> index 5763e64..7a2a90f 100644
>>> --- a/net/sunrpc/svcsock.c
>>> +++ b/net/sunrpc/svcsock.c
>>> @@ -345,7 +345,6 @@ static void svc_sock_setbufsize(struct socket *sock, unsigned int snd,
>>>      lock_sock(sock->sk);
>>>      sock->sk->sk_sndbuf = snd * 2;
>>>      sock->sk->sk_rcvbuf = rcv * 2;
>>> -    sock->sk->sk_userlocks |= SOCK_SNDBUF_LOCK|SOCK_RCVBUF_LOCK;
>>>      release_sock(sock->sk);
>>>  #endif
>>>  }
>>> @@ -797,23 +796,6 @@ static int svc_tcp_recvfrom(struct svc_rqst *rqstp)
>>>              test_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags),
>>>              test_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags));
>>>
>>> -    if (test_and_clear_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags))
>>> -            /* sndbuf needs to have room for one request
>>> -             * per thread, otherwise we can stall even when the
>>> -             * network isn't a bottleneck.
>>> -             *
>>> -             * We count all threads rather than threads in a
>>> -             * particular pool, which provides an upper bound
>>> -             * on the number of threads which will access the socket.
>>> -             *
>>> -             * rcvbuf just needs to be able to hold a few requests.
>>> -             * Normally they will be removed from the queue
>>> -             * as soon a a complete request arrives.
>>> -             */
>>> -            svc_sock_setbufsize(svsk->sk_sock,
>>> -                                (serv->sv_nrthreads+3) * serv->sv_max_mesg,
>>> -                                3 * serv->sv_max_mesg);
>>> -
>>>      clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
>>>
>>>      /* Receive data. If we haven't got the record length yet, get
>>> @@ -1061,15 +1043,6 @@ static void svc_tcp_init(struct svc_sock *svsk, struct svc_serv *serv)
>>>
>>>              tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF;
>>>
>>> -            /* initialise setting must have enough space to
>>> -             * receive and respond to one request.
>>> -             * svc_tcp_recvfrom will re-adjust if necessary
>>> -             */
>>> -            svc_sock_setbufsize(svsk->sk_sock,
>>> -                                3 * svsk->sk_xprt.xpt_server->sv_max_mesg,
>>> -                                3 * svsk->sk_xprt.xpt_server->sv_max_mesg);
>>> -
>>> -            set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
>>>              set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
>>>              if (sk->sk_state != TCP_ESTABLISHED)
>>>                      set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
>>> @@ -1140,8 +1113,14 @@ static struct svc_sock *svc_setup_socket(struct svc_serv *serv,
>>>      /* Initialize the socket */
>>>      if (sock->type == SOCK_DGRAM)
>>>              svc_udp_init(svsk, serv);
>>> -    else
>>> +    else {
>>> +            /* initialise setting must have enough space to
>>> +             * receive and respond to one request.
>>> +             */
>>> +            svc_sock_setbufsize(svsk->sk_sock, 4 * serv->sv_max_mesg,
>>> +                                    4 * serv->sv_max_mesg);
>>>              svc_tcp_init(svsk, serv);
>>> +    }
>>>
>>>      /*
>>>       * We start one listener per sv_serv.  We want AF_INET
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
>> the body of a message to majordomo@vger.kernel.org
>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>> Please read the FAQ at  http://www.tux.org/lkml/
>
>