|
Home > Archive > Apache Directory Project > August 2005 > Performance of ApacheDS versus OpenLdap. First results!
You are viewing an archived Text-only version of the thread.
To view this thread in it's original format and/or if you want to reply to
this thread please [click here]
| Author |
Performance of ApacheDS versus OpenLdap. First results!
|
|
| Emmanuel Lecharny 2005-08-29, 5:45 pm |
| Thanks to JMeter !
1
thread,
launched 1000 times. Bind, Search and UnBind
ApacheDS
CPU=
100%
Request
#sample
Average
Median
90%
line
Min
Max
Error %
Throughput
KB/sec
Bind
1000
4
2
3
1
191
0.00%
41.1/sec
8,06
Search
1000
3
2
3
1
184
0.00%
41.1/sec
14,6
UnBind
1000
0
1
1
0
185
0.00%
41.1/sec
8,34
TOTAL
3000
2
2
2
0
191
0.00%
123.2/sec
30,99
OpenLdap
CPU=100%
Request
#sample
Average
Median
90%
line
Min
Max
Error %
Throughput
KB/sec
Bind
1000
4
2
3
1
211
0.00%
47.6/sec
9,63
Search
1000
1
1
2
0
182
0.00%
47.6/sec
16,79
UnBind
1000
0
0
2
0
7
0.00%
47.6/sec
9,68
TOTAL
3000
2
1
3
0
211
0.00%
142.9/sec
36,09
5
threads, launched 200 times. Bind, Search and UnBind
ApacheDS
CPU=
100%
Request
#sample
Average
Median
90%
line
Min
Max
Error %
Throughput
KB/sec
Bind
1000
65
29
218
2
777
0.00%
57.7/sec
11,32
Search
1000
9
6
14
1
339
0.00%
57.7/sec
20,51
UnBind
1000
1
1
3
0
34
0.00%
57.7/sec
11,72
TOTAL
3000
25
6
53
0
777
0.00%
172.9/sec
43,51
OpenLdap
CPU=100%
Request
#sample
Average
Median
90%
line
Min
Max
Error %
Throughput
KB/sec
Bind
1000
58
28
206
13
244
0.00%
53.2/sec
10,76
Search
1000
25
11
18
3
229
0.00%
52.7/sec
18,59
UnBind
1000
0
0
2
0
184
0.00%
52.8/sec
10,72
TOTAL
3000
28
11
31
0
244
0.00%
158/sec
39,92
| |
| David Boreham 2005-08-29, 5:45 pm |
|
I have a suspicion that your figures may not be right.
Is it possible that you're actually measuring the performance of the
test client ?
I know that I've seen nearly 10,000 searches/s from OL.
It's a little strange that the ApacheDS and OL numbers are almost
exactly the same too.
Also strange that binds and unbinds are done with exactly
the same throughput as searches.
Perhaps I'm not reading the table correctly : I'm looking
at the 'throughput' column.
| |
| Emmanuel Lecharny 2005-08-29, 5:45 pm |
| On Mon, 2005-08-29 at 15:03 -0600, David Boreham wrote:
> I have a suspicion that your figures may not be right.
>
> Is it possible that you're actually measuring the performance of the
> test client ?
I'm using Jmeter to do the tests. It really sends requests to the
server, I bet.
>
> I know that I've seen nearly 10,000 searches/s from OL.
My computer is a little bit loaded at the moment ;) And as I'm just
doing a very simplistic bind/search/unbind request, it's not really a
search test.
>
> It's a little strange that the ApacheDS and OL numbers are almost
> exactly the same too.
you mean apacheds ~ openldap? or that we have the same values in the
Throuhput column?
>
> Also strange that binds and unbinds are done with exactly
> the same throughput as searches.
This is just plain normal. As they are sequential, what is counted is
the throughput of the sequence B-S-U. The Total number is worthless, I
think.
>
> Perhaps I'm not reading the table correctly : I'm looking
> at the 'throughput' column.
This is the good column, but just forget about the Total Throughput.
However, do NOT take this test as something serious ;) This is just a
little test I did to help us keeping the good pace. We have a pretty
fast Ldap Server, which need a LOT of work to be production-ready.
Sometime we also need some good marks ;)
| |
| David Boreham 2005-08-29, 5:45 pm |
|
>
>This is just plain normal. As they are sequential, what is counted is
>the throughput of the sequence B-S-U. The Total number is worthless, I
>think.
>
>
Ah, perhaps now I understand. It may be that you are measuring the
performance of the OS'es TCP implementation rather than the LDAP servers ;)
If the test makes a connection, then performs bind/search/unbind (which
will disconnect), then the overwhelming proportion of the effort to
handle the load
goes in setting up and tearing down the TCP connection. This would also
explain why the figures are similar.
Did you monitor the proportion of kernel to user time on the test machine ?
High kernel time (>15%) would tend to indicate an I/O-limited test.
| |
| Emmanuel Lecharny 2005-08-29, 5:45 pm |
| On Mon, 2005-08-29 at 15:26 -0600, David Boreham wrote:
>
> Ah, perhaps now I understand. It may be that you are measuring the
> performance of the OS'es TCP implementation rather than the LDAP
> servers ;)
Ooops, yes, you are right ! I always test the TCP stack through a
limited test of ApacheDS (and sometime I do the same with
Databases) ! ;-] eh eh eh
>
> If the test makes a connection, then performs bind/search/unbind
> (which
> will disconnect), then the overwhelming proportion of the effort to
> handle the load
> goes in setting up and tearing down the TCP connection. This would
> also
> explain why the figures are similar.
>
> Did you monitor the proportion of kernel to user time on the test
> machine ?
> High kernel time (>15%) would tend to indicate an I/O-limited test.
6% system. Which is kind of high...
Ok, ok. Do not push this test too far. It's not realistic. It's just a
kind of baby test (very very young ! It's just one cell, actually, and
it will divide soon to two cells, then 4, then ...)
Call it very premature ;)
Nevertheless, I like those results !
Emmanuel
|
|
|
|
|