In measuring the bandwidth across gigabit ethernet between 2 Alphas
running Redhat 4.2, we've noticed some quirks for certain message sizes.
When the message size is roughly half of the TCP buffer size, the
performance drops by a considerable amount, only to rebound at around
75% of the buffer size (see the measurements below). Can anyone
explain to me what Linux could be doing to cause this problem? This is
a serious shortcoming that we've observed on both Alphas and Pentium Pros
running Linux, while FreeBSD shows no similar problems on the same
hardware. In applications, I've seen intermittent delays of upto 200 ms
for messages in this range, while breaking up the same messages into
multiple packets avoids these problems.
I'ld also like to increase the maximum TCP buffer size beyond the
current maximum of 65535 bytes. How easy is this, and how do I go
about doing it?
Dave Turner
NetPIPE results for a 20 kByte buffer size
6: 7168 bytes 783 times --> 159.16 Mbps in 0.000344 sec
7: 8189 bytes 726 times --> 160.72 Mbps in 0.000389 sec
8: 8192 bytes 642 times --> 160.81 Mbps in 0.000389 sec
9: 8195 bytes 642 times --> 160.87 Mbps in 0.000389 sec
10: 9213 bytes 642 times --> 165.33 Mbps in 0.000425 sec
11: 9216 bytes 587 times --> 59.48 Mbps in 0.001182 sec
12: 9219 bytes 211 times --> 60.19 Mbps in 0.001169 sec
13: 10237 bytes 213 times --> 71.13 Mbps in 0.001098 sec
14: 10240 bytes 227 times --> 69.98 Mbps in 0.001116 sec
15: 10243 bytes 223 times --> 68.64 Mbps in 0.001139 sec
16: 11261 bytes 219 times --> 75.85 Mbps in 0.001133 sec
17: 11264 bytes 220 times --> 75.77 Mbps in 0.001134 sec
18: 11267 bytes 220 times --> 76.69 Mbps in 0.001121 sec
19: 12285 bytes 222 times --> 96.63 Mbps in 0.000970 sec
20: 12288 bytes 257 times --> 97.71 Mbps in 0.000959 sec
21: 12291 bytes 259 times --> 95.47 Mbps in 0.000982 sec
22: 13309 bytes 253 times --> 69.96 Mbps in 0.001451 sec
23: 13312 bytes 171 times --> 70.71 Mbps in 0.001436 sec
24: 13315 bytes 173 times --> 70.85 Mbps in 0.001434 sec
25: 14333 bytes 173 times --> 87.66 Mbps in 0.001248 sec
26: 14336 bytes 199 times --> 60.08 Mbps in 0.001821 sec
27: 14339 bytes 136 times --> 69.57 Mbps in 0.001573 sec
28: 15357 bytes 158 times --> 93.38 Mbps in 0.001255 sec
29: 15360 bytes 198 times --> 99.41 Mbps in 0.001179 sec
30: 15363 bytes 211 times --> 100.50 Mbps in 0.001166 sec
31: 16381 bytes 212 times --> 151.15 Mbps in 0.000827 sec
32: 16384 bytes 300 times --> 151.78 Mbps in 0.000824 sec
33: 16387 bytes 301 times --> 151.41 Mbps in 0.000826 sec
34: 17405 bytes 300 times --> 153.07 Mbps in 0.000868 sec
35: 17408 bytes 286 times --> 154.35 Mbps in 0.000860 sec
36: 17411 bytes 288 times --> 153.93 Mbps in 0.000863 sec
*******************************************************************
* Dr Dave Turner Asst Scientist - Ames Lab - ISU *
* 524 Physics turner@ameslab.gov *
* Iowa State University Phone:(515)-294-8872 *
* Ames, IA 50011 Fax: (515)-294-0689 *
*******************************************************************
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu