• Eric Dumazet's avatar
    tcp: refine TSO splits · d4589926
    Eric Dumazet authored
    
    
    While investigating performance problems on small RPC workloads,
    I noticed linux TCP stack was always splitting the last TSO skb
    into two parts (skbs). One being a multiple of MSS, and a small one
    with the Push flag. This split is done even if TCP_NODELAY is set,
    or if no small packet is in flight.
    
    Example with request/response of 4K/4K
    
    IP A > B: . ack 68432 win 2783 <nop,nop,timestamp 6524593 6525001>
    IP A > B: . 65537:68433(2896) ack 69632 win 2783 <nop,nop,timestamp 6524593 6525001>
    IP A > B: P 68433:69633(1200) ack 69632 win 2783 <nop,nop,timestamp 6524593 6525001>
    IP B > A: . ack 68433 win 2768 <nop,nop,timestamp 6525001 6524593>
    IP B > A: . 69632:72528(2896) ack 69633 win 2768 <nop,nop,timestamp 6525001 6524593>
    IP B > A: P 72528:73728(1200) ack 69633 win 2768 <nop,nop,timestamp 6525001 6524593>
    IP A > B: . ack 72528 win 2783 <nop,nop,timestamp 6524593 6525001>
    IP A > B: . 69633:72529(2896) ack 73728 win 2783 <nop,nop,timestamp 6524593 6525001>
    IP A > B: P 72529:73729(1200) ack 73728 win 2783 <nop,nop,timestamp 6524593 6525001>
    
    We can avoid this split by including the Nagle tests at the right place.
    
    Note : If some NIC had trouble sending TSO packets with a partial
    last segment, we would have hit the problem in GRO/forwarding workload already.
    
    tcp_minshall_update() is moved to tcp_output.c and is updated as we might
    feed a TSO packet with a partial last segment.
    
    This patch tremendously improves performance, as the traffic now looks
    like :
    
    IP A > B: . ack 98304 win 2783 <nop,nop,timestamp 6834277 6834685>
    IP A > B: P 94209:98305(4096) ack 98304 win 2783 <nop,nop,timestamp 6834277 6834685>
    IP B > A: . ack 98305 win 2768 <nop,nop,timestamp 6834686 6834277>
    IP B > A: P 98304:102400(4096) ack 98305 win 2768 <nop,nop,timestamp 6834686 6834277>
    IP A > B: . ack 102400 win 2783 <nop,nop,timestamp 6834279 6834686>
    IP A > B: P 98305:102401(4096) ack 102400 win 2783 <nop,nop,timestamp 6834279 6834686>
    IP B > A: . ack 102401 win 2768 <nop,nop,timestamp 6834687 6834279>
    IP B > A: P 102400:106496(4096) ack 102401 win 2768 <nop,nop,timestamp 6834687 6834279>
    IP A > B: . ack 106496 win 2783 <nop,nop,timestamp 6834280 6834687>
    IP A > B: P 102401:106497(4096) ack 106496 win 2783 <nop,nop,timestamp 6834280 6834687>
    IP B > A: . ack 106497 win 2768 <nop,nop,timestamp 6834688 6834280>
    IP B > A: P 106496:110592(4096) ack 106497 win 2768 <nop,nop,timestamp 6834688 6834280>
    
    Before :
    
    lpq83:~# nstat >/dev/null;perf stat ./super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K
    280774
    
     Performance counter stats for './super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K':
    
         205719.049006 task-clock                #    9.278 CPUs utilized
             8,449,968 context-switches          #    0.041 M/sec
             1,935,997 CPU-migrations            #    0.009 M/sec
               160,541 page-faults               #    0.780 K/sec
       548,478,722,290 cycles                    #    2.666 GHz                     [83.20%]
       455,240,670,857 stalled-cycles-frontend   #   83.00% frontend cycles idle    [83.48%]
       272,881,454,275 stalled-cycles-backend    #   49.75% backend  cycles idle    [66.73%]
       166,091,460,030 instructions              #    0.30  insns per cycle
                                                 #    2.74  stalled cycles per insn [83.39%]
        29,150,229,399 branches                  #  141.699 M/sec                   [83.30%]
         1,943,814,026 branch-misses             #    6.67% of all branches         [83.32%]
    
          22.173517844 seconds time elapsed
    
    lpq83:~# nstat | egrep "IpOutRequests|IpExtOutOctets"
    IpOutRequests                   16851063           0.0
    IpExtOutOctets                  23878580777        0.0
    
    After patch :
    
    lpq83:~# nstat >/dev/null;perf stat ./super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K
    280877
    
     Performance counter stats for './super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K':
    
         107496.071918 task-clock                #    4.847 CPUs utilized
             5,635,458 context-switches          #    0.052 M/sec
             1,374,707 CPU-migrations            #    0.013 M/sec
               160,920 page-faults               #    0.001 M/sec
       281,500,010,924 cycles                    #    2.619 GHz                     [83.28%]
       228,865,069,307 stalled-cycles-frontend   #   81.30% frontend cycles idle    [83.38%]
       142,462,742,658 stalled-cycles-backend    #   50.61% backend  cycles idle    [66.81%]
        95,227,712,566 instructions              #    0.34  insns per cycle
                                                 #    2.40  stalled cycles per insn [83.43%]
        16,209,868,171 branches                  #  150.795 M/sec                   [83.20%]
           874,252,952 branch-misses             #    5.39% of all branches         [83.37%]
    
          22.175821286 seconds time elapsed
    
    lpq83:~# nstat | egrep "IpOutRequests|IpExtOutOctets"
    IpOutRequests                   11239428           0.0
    IpExtOutOctets                  23595191035        0.0
    
    Indeed, the occupancy of tx skbs (IpExtOutOctets/IpOutRequests) is higher :
    2099 instead of 1417, thus helping GRO to be more efficient when using FQ packet
    scheduler.
    
    Many thanks to Neal for review and ideas.
    Signed-off-by: default avatarEric Dumazet <edumazet@google.com>
    Cc: Yuchung Cheng <ycheng@google.com>
    Cc: Neal Cardwell <ncardwell@google.com>
    Cc: Nandita Dukkipati <nanditad@google.com>
    Cc: Van Jacobson <vanj@google.com>
    Acked-by: default avatarNeal Cardwell <ncardwell@google.com>
    Tested-by: default avatarNeal Cardwell <ncardwell@google.com>
    Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
    d4589926
tcp_output.c 91.8 KB