You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@httpd.apache.org by Dean Gaudet <dg...@arctic.org> on 1997/08/24 01:36:26 UTC
this is depressing
If I run the code with perform_idle_server_maintenance doing nothing,
zilch, that is we don't spawn anything more than what we started with.
Then I can get 422 r/s versus 409 r/s for the default code. I'm attacking
it with 30 simultaneous clients ... maybe zeusbench just sucks bad. Even
the latency doesn't get worse. I need a better test environment.
Dean
# default code (well plus some of my development stuff)
[33]benchlark:~% zb 127.0.0.1 /6k -t 30 -c 30 -p 8080
---
Server: Apache/1.3a2-dev
Doucment Length: 6144
Concurency Level: 30
Time taken for tests: 30.005 seconds
Complete requests: 12291
Failed requests: 0
Bytes transfered: 78611079
HTML transfered: 75595864
Requests per seconds: 409.63
Transfer rate: 2619.93 kb/s
Connnection Times (ms)
min avg max
Connect: 0 8 38
Total: 25 71 171
---
# code without any scoreboard maintenance
[34]benchlark:~% zb 127.0.0.1 /6k -t 30 -c 30 -p 8080
---
Server: Apache/1.3a2-dev
Doucment Length: 6144
Concurency Level: 30
Time taken for tests: 30.001 seconds
Complete requests: 12669
Failed requests: 0
Bytes transfered: 80967307
HTML transfered: 77862177
Requests per seconds: 422.29
Transfer rate: 2698.82 kb/s
Connnection Times (ms)
min avg max
Connect: 0 2 12
Total: 23 68 114
---
Re: this is depressing
Posted by Dean Gaudet <dg...@arctic.org>.
On Sat, 23 Aug 1997, Marc Slemko wrote:
> Odd. What sort of scoreboard are you using?
shmget(). I'm running with very few changes right now, just a fix for the
last_rtime initialization, and buffered logs. I've had to crank
MaxRequestsPerChild up to 999999 because the server w/o scoreboard
maintenance eventually stops serving hits. The two servers are the same
except for the scoreboard maintenance (or lack thereof).
> Remember, zeusbench uses a single process. As long as the server isn't
> overloaded in the wrong way, responses will be quick enough that the other
> requests will be processed almost right away. Since zb can't read
> multiple responses at exactly the same time anyway...
Aw shoot yeah I'm going to run httpd at nice +4 and zb at nice +0 ...
otherwise the server overloads the benchmark. Some day I'll try this at
the office where I can probably get a half-dozen ppros on 100baseT
involved.
Here's 5 minute results. It's worse.
Dean
# with no scoreboard maintenance
[44]benchlark:~% zb 127.0.0.1 /6k -t 300 -n 150000 -c 30 -p 8080
---
Server: Apache/1.3a2-dev
Doucment Length: 6144
Concurency Level: 30
Time taken for tests: 300.005 seconds
Complete requests: 125522
Failed requests: 0
Bytes transfered: 801972836
HTML transfered: 771219456
Requests per seconds: 418.40
Transfer rate: 2673.20 kb/s
Connnection Times (ms)
min avg max
Connect: 0 3 14
Total: 21 69 153
---
# with 5s scoreboard maintenance
[45]benchlark:~% zb 127.0.0.1 /6k -t 300 -n 150000 -c 30 -p 8080
---
Server: Apache/1.3a2-dev
Doucment Length: 6144
Concurency Level: 30
Time taken for tests: 300.020 seconds
Complete requests: 117182
Failed requests: 0
Bytes transfered: 748723147
HTML transfered: 720010862
Requests per seconds: 390.58
Transfer rate: 2495.58 kb/s
Connnection Times (ms)
min avg max
Connect: 0 16 39
Total: 24 75 231
---
Re: this is depressing
Posted by Marc Slemko <ma...@worldgate.com>.
On Sat, 23 Aug 1997, Dean Gaudet wrote:
> Oh sorry I forgot to mention that my dev version is running at 5s right
> now ... with the default 1s we get around 390 r/s.
Odd. What sort of scoreboard are you using?
>
> I'm depressed that for some reason we're able to satisfy 30 clients with
> only 5 children, and doing scoreboard maintenance is so expensive. But
> this is quite easily because my test setup sucks.
Remember, zeusbench uses a single process. As long as the server isn't
overloaded in the wrong way, responses will be quick enough that the other
requests will be processed almost right away. Since zb can't read
multiple responses at exactly the same time anyway...
What if you start with 30 children?
>
> It'd probably be different if I ran a 10 minute trial... in 30 seconds we
> pay for the fork()s. I'll try that.
>
> Dean
>
> On Sat, 23 Aug 1997, Marc Slemko wrote:
>
> > Are you saying you are suprised by how much it does change or by how much
> > it doesn't change?
> >
> > I am suprised by how big the improvement without the idle server maint.
> > is. What happens if you change SCOREBOARD_MAINTENANCE_INTERVAL to 5s or
> > something?
>
Re: this is depressing
Posted by Dean Gaudet <dg...@arctic.org>.
Oh sorry I forgot to mention that my dev version is running at 5s right
now ... with the default 1s we get around 390 r/s.
I'm depressed that for some reason we're able to satisfy 30 clients with
only 5 children, and doing scoreboard maintenance is so expensive. But
this is quite easily because my test setup sucks.
It'd probably be different if I ran a 10 minute trial... in 30 seconds we
pay for the fork()s. I'll try that.
Dean
On Sat, 23 Aug 1997, Marc Slemko wrote:
> Are you saying you are suprised by how much it does change or by how much
> it doesn't change?
>
> I am suprised by how big the improvement without the idle server maint.
> is. What happens if you change SCOREBOARD_MAINTENANCE_INTERVAL to 5s or
> something?
Re: this is depressing
Posted by Marc Slemko <ma...@worldgate.com>.
Are you saying you are suprised by how much it does change or by how much
it doesn't change?
I am suprised by how big the improvement without the idle server maint.
is. What happens if you change SCOREBOARD_MAINTENANCE_INTERVAL to 5s or
something?
On Sat, 23 Aug 1997, Dean Gaudet wrote:
> If I run the code with perform_idle_server_maintenance doing nothing,
> zilch, that is we don't spawn anything more than what we started with.
> Then I can get 422 r/s versus 409 r/s for the default code. I'm attacking
> it with 30 simultaneous clients ... maybe zeusbench just sucks bad. Even
> the latency doesn't get worse. I need a better test environment.
>
> Dean
>
> # default code (well plus some of my development stuff)
> [33]benchlark:~% zb 127.0.0.1 /6k -t 30 -c 30 -p 8080
>
> ---
> Server: Apache/1.3a2-dev
> Doucment Length: 6144
> Concurency Level: 30
> Time taken for tests: 30.005 seconds
> Complete requests: 12291
> Failed requests: 0
> Bytes transfered: 78611079
> HTML transfered: 75595864
> Requests per seconds: 409.63
> Transfer rate: 2619.93 kb/s
>
> Connnection Times (ms)
> min avg max
> Connect: 0 8 38
> Total: 25 71 171
> ---
>
> # code without any scoreboard maintenance
> [34]benchlark:~% zb 127.0.0.1 /6k -t 30 -c 30 -p 8080
>
> ---
> Server: Apache/1.3a2-dev
> Doucment Length: 6144
> Concurency Level: 30
> Time taken for tests: 30.001 seconds
> Complete requests: 12669
> Failed requests: 0
> Bytes transfered: 80967307
> HTML transfered: 77862177
> Requests per seconds: 422.29
> Transfer rate: 2698.82 kb/s
>
> Connnection Times (ms)
> min avg max
> Connect: 0 2 12
> Total: 23 68 114
> ---
>
>