[rds-devel] Re: Fwd: Re: bcopy congestion / flow control... ?

Olaf Kirch olaf.kirch at oracle.com
Thu Feb 7 08:30:11 PST 2008


On Thursday 07 February 2008 17:26, Olaf Kirch wrote:
> I first need to do some performance testing to
> see if they actually change anything for the better.

Here's some data with -c -t16 -d16 -q2k -T10, with and without
the congestion monitor patch. The cong_send_error is from
rds-info -c output. As you can see, the patch almost
eliminates any sendmsg() calls to a blocked destination.

With:
tsks   tx/s  tx+rx K/s  rw+rr K/s tx us/c   rtt us cpu %
  16  57165  234664.50       0.00   11.03  6995.10 88.61
  16  56523  231846.82       0.00   13.90  6850.03 87.06
  16  57329  235077.69       0.00   14.26  6928.14 88.27
  16  56644  232725.80       0.00   13.27  6818.35 87.93
  16  53708  220394.22       0.00    9.75  7566.85 85.12
  16  56534  232124.08       0.00   13.29  7115.16 89.74
  16  56458  231775.76       0.00   10.97  6992.60 88.61
  16  62094  234593.76       0.00   11.94  6055.03 78.22
  16  55710  228850.61       0.00   12.59  7343.49 90.86
  16  40387  165520.50       0.00   15.24  6901.72 74.94
---------------------------------------------
  16  54709  224838.12       0.00   12.69  7013.30 85.87  (average)
          cong_send_error               61


Without:
  16  53084  215330.30       0.00   12.05 17275.83 98.30
  16  52496  214284.69       0.00   10.97 16090.59 97.98
  16  52836  215839.99       0.00   10.84 15903.95 96.14
  16  52780  215447.82       0.00   17.88 13803.93 96.02
  16  54858  223465.00       0.00   10.77 13802.60 95.04
  16  53378  218084.21       0.00   10.22 14450.08 94.39
  16  52483  213937.36       0.00   10.67 13866.91 95.70
  16  53911  220019.19       0.00   10.65 14293.49 95.03
  16  54424  222617.79       0.00   15.82 13304.42 93.24
  16  46905  191467.42       0.00   13.78 13508.08 88.06
---------------------------------------------
  16  52655  215050.01       0.00   12.36 14652.36 94.99  (average)
          cong_send_error          1359177

Performance doesn't change a lot, which is kind of expected,
but RTT and CPU utilization improve noticeably.

At 10 microseconds per sendmsg call, cutting out 1.3 million
sendmsg calls amounts to 13 seconds of CPU time.

Olaf
-- 
Olaf Kirch  |  --- o --- Nous sommes du soleil we love when we play
okir at lst.de |    / | \   sol.dhoop.naytheet.ah kin.ir.samse.qurax



More information about the rds-devel mailing list