directory-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Emmanuel Lecharny <elecha...@gmail.com>
Subject Re: Benchmarks feedback
Date Tue, 27 Mar 2007 00:03:50 GMT
ole ersoy a écrit :

> I wonder if disk access could be the bottle neck?

I have done some more test, after I realized that I didn't correctly set 
the cache siez, nor th ememory dedicated to the JVM : Tha cache was 
containing only 100 entries, so each read was done using the File System 
(but with 20 Gb ram, a read is done entirely in memory).

I get better results : up to 3000 req/s on the random searches, which is 
twice better than what I observed before. Cache helps !

Now, I get some run queues (vmstat 1):
procs -----------memory---------- ---swap-- -----io---- -system-- 
----cpu----
 r  b   swpd   free   buff  cache   si   so    bi    bo   in   cs us sy 
id wa
10  0      0 13379076 361024 5123816    0    0     0    44 14314 32902 
15  4 80  0
 6  0      0 13378356 361024 5123816    0    0     0     0 14487 40173 
20  4 75  0
12  0      0 13378404 361024 5123816    0    0     0     0 13614 39101 
15  4 81  0
18  0      0 13377824 361024 5123816    0    0     0     0 14611 41310 
16  4 80  0
 3  0      0 13377848 361024 5123816    0    0     0     0 13470 46931 
18  3 78  0
 0  0      0 13376148 361024 5123816    0    0     0     0 6836 24492 
15  2 83  0
 7  0      0 13391344 361024 5123752    0    0     0    72 2081 5066 14  
4 82  0
11  0      0 13383120 361024 5123784    0    0     0     0 17499 61286 
18  3 79  0

The clients are sometime burning some heavy cpu cycles :

  PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND
24687 elecharn  20   0  798m 294m  13m S   93  1.5  68:50.31 java
 3835 elecharn  17   0 1102m  82m 1492 S   57  0.4   0:01.71 java
 3831 elecharn  23   0 1102m  82m 1492 S   54  0.4   0:01.64 java
 3833 elecharn  17   0 1103m  82m 1492 S   52  0.4   0:01.58 java
 3837 elecharn  17   0 1103m  81m 1492 S   47  0.4   0:01.42 java

I guess that I need to run more injectors with less threads

>
> Would be interesting to see what happens if we get
> a memory resident partition and run the same test against that.

Well, with a correctly sized cache, this is what we obtain : 3000 req/s, 
at best (average around 2000 req/s)

>
> Would be really sweet to get the whole box humming.  Thanks Matt and
> Emmanuel!
> It's great to be aware of this type of stuff.  I have a couple of 20 cpu
> servers in my closet,
> but  I have not had a chance to dust them off yet (Yeah, umhummm, sure).

I also have a PDP 11 somwhere, but it's currently running since last 
year computing 45! ...

Emmanuel


Mime
View raw message