* Route hash code comparison
@ 2003-05-20 16:12 Robert Olsson
2003-05-21 2:42 ` David S. Miller
0 siblings, 1 reply; 2+ messages in thread
From: Robert Olsson @ 2003-05-20 16:12 UTC (permalink / raw)
To: netdev; +Cc: Robert.Olsson
Hello!
Route hash code comparison on fairly loaded router at Uppsala Universitet.
Router is loaded both in terms of packet load but also in IP-flows from
(tens) of thousands of users.
The hash table was modified to count linear search in the cache which was
printed with rtstat (in_search). Router has 32 k hash buckets which is where
the GC starts. This point is choosen for hash algorithm comparison as well.
Old hash:
107694/41990 ==> 2.6 linear searches for a packet look-up at 32 k entries
New hash:
33334/38927 ==> 0.86 linear searches for a packet look-up at 32 k entries
The new hash have random keys but seems to be pretty consistent in term of
hashing at least from observerations. No DoS experiments done.
[rstat: Table is edited]
size == route cache size
hit == IN: total number of cache hits per sec
tot == IN: total number of cache misses per sec
GC == GC: garbage collection calls per sec
in_search == HASH: input hash list search per sec
Linux 2.5.66
size IN: hit tot GC in_search
==============================================
1467 39622 1255 0 80331
7034 38195 268 0 19382
9743 51429 129 0 32903
11633 50069 91 0 40124
13316 48456 83 0 41458
14846 45883 75 0 47253
16233 44672 65 0 52647
17574 43786 65 0 57167
18855 43227 61 0 61784
20106 43341 59 0 65187
21316 43389 57 0 68505
22481 43386 55 0 73778
23694 42748 58 0 77601
24819 42549 54 0 83328
26180 41474 65 0 82393
27331 41754 53 0 90068
28408 41626 50 0 94326
29415 40731 47 0 97407
30424 42193 47 0 102571
31382 41313 46 0 102311
32348 41990 46 0 107694 * Analys1
33524 42104 57 377 108611
34527 42170 48 501 118257
34148 44107 45 468 120707
35196 43345 50 524 116961
36064 42004 42 434 119295
35868 41805 49 504 119121
36838 42299 45 487 116961
Linux 2.5.66 with hash code from 2.5.69.
size IN: hit tot GC in_search
==============================================
6031 34776 0 0 7150
8517 38168 119 0 5455
10395 38431 86 0 6501
11954 37002 73 0 8972
13546 37776 74 0 10742
14948 38595 65 0 12585
16406 37981 67 0 13827
17802 38479 64 0 15437
18984 38392 54 0 17684
20155 37989 54 0 18733
21210 39508 48 0 20084
22466 40834 57 0 21299
23622 39226 52 0 23221
24836 40965 56 0 24183
26157 39683 63 0 24764
27291 40309 54 0 27793
28417 38468 54 0 25867
29358 39199 44 0 29093
30302 39159 45 0 29722
31477 38284 56 0 31246
32449 38927 46 0 33334 * Analys2
33386 39031 44 308 31837
34300 39432 42 457 35057
34117 38106 48 522 33129
35155 37887 48 519 32823
36187 37222 48 516 31986
36326 37125 50 539 33914
37257 38814 44 465 35457
Cheers.
--ro
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: Route hash code comparison
2003-05-20 16:12 Route hash code comparison Robert Olsson
@ 2003-05-21 2:42 ` David S. Miller
0 siblings, 0 replies; 2+ messages in thread
From: David S. Miller @ 2003-05-21 2:42 UTC (permalink / raw)
To: Robert.Olsson; +Cc: netdev
From: Robert Olsson <Robert.Olsson@data.slu.se>
Date: Tue, 20 May 2003 18:12:57 +0200
The new hash have random keys but seems to be pretty consistent in
term of hashing at least from observerations. No DoS experiments
done.
Thank you for posting these test results Robert.
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2003-05-21 2:42 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2003-05-20 16:12 Route hash code comparison Robert Olsson
2003-05-21 2:42 ` David S. Miller
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).