You are viewing a plain text version of this content. The canonical link for it is here.
Posted to server-dev@james.apache.org by "Noel J. Bergman" <no...@devtech.com> on 2002/10/16 10:08:58 UTC
Additional Load Test Data
This summarizes testing data for the two experimental builds. The numbers
reported here are with JDK 1.3.1 Hotspot Server. I have attached the raw
data log, which also includes the same tests using Hotspot Client.
Interestingly, the behavior is almost identical; except for a significant
initial hit when first starting to use the James classes.
./postal -m 1 -p 1 -c 50000 amitower userlist - | tee -a test.log
time,messages,data(K),errors,connections,SSL connections
Memory %P Scheduler Build Watchdog Build %P Memory
--------------------------------------------------------------------
129388 -- Before Phoenix -- -- Before Phoenix -- 147768
112948 -- Before Testing -- -- Before Testing -- 134200
107764 48 00:41,1944,1669,0,1,0 01:01,1814,1556,0,1,0 43 125048
103972 58 50 124860
99736 18 00:42,2024,1741,0,0,0 01:02,1987,1710,0,0,0 26 124860
95420 20 19 119172
91148 25 00:43,2049,1746,0,0,0 01:03,2051,1755,0,0,0 15 115208
86804 22 14 111048
82424 23 00:44,2061,1753,0,0,0 01:04,2057,1749,0,0,0 14 106844
78048 25 14 102572
73672 24 00:45,2066,1747,0,0,0 01:05,2048,1742,0,0,0 15 98224
69296 25 14 93876
64948 24 00:46,2062,1743,0,0,0 01:06,2053,1748,0,0,0 14 89528
60548 24 14 85180
56172 25 00:47,2065,1763,0,0,0 01:07,2055,1753,0,0,0 14 80812
51796 24 14 76468
47424 23 00:48,2062,1751,0,0,0 01:08,2068,1795,0,0,0 14 72136
43056 24 14 67748
38656 24 00:49,2069,1760,0,0,0 01:09,2056,1758,0,0,0 15 63400
34292 24 14 59032
29596 23 00:50,2077,1773,0,0,0 01:10,2042,1738,0,0,0 15 54624
25212 24 13 50256
20784 24 00:51,2086,1813,0,0,0 01:11,2053,1771,0,0,0 15 45912
16340 23 15 41532
11952 24 00:52,2082,1808,0,0,0 01:12,2051,1768,0,0,0 14 37228
7552 24 14 32840
3408 23 00:53,2070,1786,0,1,0 01:13,2057,1764,0,1,0 14 28504
3188 28 13 24176
3116 24 00:54,2060,1753,0,0,0 01:14,2058,1775,0,0,0 12 19812
3072 22 13 15444
3188 22 00:55,2066,1760,0,0,0 01:15,2049,1754,0,0,0 12 14388
Basically, we're losing roughly 8.5K - 9K per 2050 1K messages. The test
uses a single connection which never closes (I push 50,000 1K messages
through it, and it doesn't run to completion before I stop the test).
I want to stress that this behavior is common to both builds and both JITs.
Getting the same results for this test as compared with the previous test
would seem to indicate that it is not dependent upon the number of
connections. The next test will use smaller numbers of large messages to
see if it is message count or traffic that is impacting us.
--- Noel
RE: Additional Load Test Data (correction and addition)
Posted by "Noel J. Bergman" <no...@devtech.com>.
> Basically, we're losing roughly 8.5K - 9K per 2050 1K messages.
Brain cramp ... but it's 4AM. Not KBytes. MBytes. Those numbers are
already in KBytes. We're losing 8.5 to 9 megabytes per 2050 1K messages.
And I can now report the following results before calling it a night:
./postal -m 5000 -p 1 -c 1000 amitower userlist - | tee -a test.log
time,messages,data(K),errors,connections,SSL connections
04:14,11,26533,0,1,0
04:15,10,26628,0,0,0
04:16,10,26966,0,0,0
04:17,13,26884,0,0,0
04:18,12,27075,0,0,0
04:19,10,27096,0,0,0
$ vmstat 30
procs memory swap io system
cpu
r b w swpd free buff cache si so bi bo in cs us sy
id
2 0 0 97532 89636 39184 26020 0 0 0 259 1413 1809 24 4
72
2 0 0 97532 87456 41148 26020 0 0 0 319 1627 1950 25 6
70
12 0 0 97532 75352 52348 26020 0 0 0 458 1888 1900 27 5
68
12 0 0 97532 70292 57028 26020 0 0 0 365 1694 1907 25 5
70
8 0 0 97532 70288 57028 26020 0 0 0 458 1889 1914 25 6
69
2 0 0 97532 70280 57028 26020 0 0 0 413 1803 1907 26 6
68
12 0 0 97532 70272 57028 26020 0 0 0 316 1610 1930 25 5
70
12 0 0 97532 70268 57028 26020 0 0 0 455 1882 1893 25 7
68
12 0 0 97532 70268 57028 26020 0 0 0 368 1714 1916 25 5
70
12 0 0 97532 70264 57028 26020 0 0 0 365 1721 1947 25 5
69
5 0 0 97532 70260 57028 26020 0 0 0 460 1902 1897 25 6
68
2 0 0 97532 70252 57028 26020 0 0 0 360 1704 1932 24 5
71
That was a quick 6 minute test with large messages, a single connection, and
relatively few messages. We are losing 4K to 8K per 30 second check, so
between 8K and 16K per 10 5000 KByte messages (I attribute the initial
larger loss to the requisite memory buffer required to hold storage). And
yes, this time I do mean KBytes, not MBytes. :-)
The JIT for that test was Hotspot Client.
--- Noel
--
To unsubscribe, e-mail: <ma...@jakarta.apache.org>
For additional commands, e-mail: <ma...@jakarta.apache.org>