[TEST/REVIEW] cpu time accounting patch, step 2
Andrew Gallatin
gallatin at cs.duke.edu
Fri Feb 10 07:22:33 PST 2006
Poul-Henning Kamp writes:
> Tests, reviews etc please.
Here are some benchmark results from my "AMD Athlon(tm) 64 X2 Dual
Core Processor 3800+" running an SMP kernel for both loopback and
10GbE TCP networks. The executive summary is that there's a 41%
improvement in loopback pingpong, and a 12.5% improvement in a
pingpong test with a remote linux machine over 10GbE. Since the linux
machine isn't infinately fast, the improvement is probably a bit more
than 12.5% ;)
The machine seemed stable, and I haven't seen anything negative
except for a tiny slowdown in sending using TCP_SENDFILE.
Thank you so much for doing this!!
Drew
Test data:
machdep.cpu_idle_hlt=0
net.isr.direct=1
dev.myri10ge.0.skip_pio_read=1
Everything else default except where noted for 10GbE TCP_RR test.
netperf -tTCP_STREAM to localhost
x loopback_stream.before
+ loopback_stream.after
+--------------------------------------------------------------------------+
|x x x x x + + + + +|
| |______A_M____| |____M_A_____| |
+--------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 5 4100.11 4221.45 4187.95 4178.016 46.300782
+ 5 4472.37 4567.2 4510 4519.416 38.960833
Difference at 95.0% confidence
341.4 +/- 62.4045
8.17134% +/- 1.49364%
(Student's t, pooled s = 42.7885)
netperf -tTCP_RR to localhost
x loopback_rr.before
+ loopback_rr.after
+--------------------------------------------------------------------------+
|x x + ++ ++|
|MA| |__MA___||
+--------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 5 27807.53 28195.85 27873.98 27964.662 163.75044
+ 5 38590.71 40167.44 39301.22 39463.23 645.1375
Difference at 95.0% confidence
11498.6 +/- 686.411
41.1182% +/- 2.45457%
(Student's t, pooled s = 470.647)
netperf -tTCP_STREAM to linux machine via 10GbE (128KB socket buffers)
x 10GbE_stream_rx.before
+ 10GbE_stream_rx.after
+--------------------------------------------------------------------------+
| + |
|x x x x x + + + + |
| |________A__M____| |_________________A_________M_______||
+--------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 5 7655.65 7723.43 7702.66 7693.382 27.071167
+ 5 7768.29 7883.34 7863.13 7830.482 57.216535
Difference at 95.0% confidence
137.1 +/- 65.2772
1.78205% +/- 0.848484%
(Student's t, pooled s = 44.7581)
x 10GbE_stream_tx.before
+ 10GbE_stream_tx.after
+--------------------------------------------------------------------------+
|+ x x + + x + + x x|
| |_|____________M__________MA_____A___________________|___________| |
+--------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 5 4640.64 4722.05 4670.05 4679.212 36.651045
+ 5 4636.54 4709.22 4657.93 4671.276 31.004687
No difference proven at 95.0% confidence
netperf -tTCP_SENDFILE to linux machine via 10GbE (128KB socket buffers)
x 10GbE_sendfile_tx.before
+ 10GbE_sendfile_tx.after
+--------------------------------------------------------------------------+
| + + + + + x xxx x|
||_________M_____A_______________| |_________A________| |
+--------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 5 7453.74 7523.76 7492.15 7490.258 24.844414
+ 5 7328.59 7430.97 7355.29 7369.944 42.886543
Difference at 95.0% confidence
-120.314 +/- 51.1132
-1.60627% +/- 0.682395%
(Student's t, pooled s = 35.0464)
netperf -tTCP_RR to linux machine via 10GbE, after setting nic interrupt
coalescing parameters to their minimum setting (dev.myri10ge.0.intr_coal_delay=1):
x 10GbE_rr.before
+ 10GbE_rr.after
+--------------------------------------------------------------------------+
|x xx + +++++|
||_AM| |_AM_||
+--------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 5 29721.79 29929.62 29863.69 29827.82 89.80339
+ 6 33323.05 33703.95 33614.14 33556.268 137.33838
Difference at 95.0% confidence
3728.45 +/- 162.431
12.4999% +/- 0.544562%
(Student's t, pooled s = 118.588)
More information about the freebsd-current
mailing list