SMP scalabilty

SMP scalabilty

Post by Frithjof Ander » Wed, 25 Jun 2003 23:24:24



We installed a vanilla Suse 8.1 on  Dell Xeon 2.4Ghz Dual box
with 1GB Ram. The running kernel is a Linux 2.4.19-64GB-SMP #1 SMP.
Now I tested the system with rather computational intensive numerical
simmulations. If I run a single job, it takes about 1h 20min to finish.
However, when I launching two identical jobs, these jobs need now 2h!
Alternativly, I can launch the same job but runing the calculation in
two threads, using pthread. Then, the job is finished in about 1hour,
indicating a performance of ca 195% CPU time.

However, I was wandering, if anybody has any suggestions, what might
be the performace decrease from 1 running job to two jobs by about 35%.
Does that have to do with the 64GB kernel, since I do not notice this
decrease on an Athlon MP Dual box, but using an Linux 2.4.18-4GB-SMP kernel?

Any suggestings appreciated.

  Frithjof

 
 
 

SMP scalabilty

Post by Sundial Service » Thu, 26 Jun 2003 01:30:01



> We installed a vanilla Suse 8.1 on  Dell Xeon 2.4Ghz Dual box
> with 1GB Ram. The running kernel is a Linux 2.4.19-64GB-SMP #1 SMP.
> Now I tested the system with rather computational intensive numerical
> simmulations. If I run a single job, it takes about 1h 20min to finish.
> However, when I launching two identical jobs, these jobs need now 2h!
> Alternativly, I can launch the same job but runing the calculation in
> two threads, using pthread. Then, the job is finished in about 1hour,
> indicating a performance of ca 195% CPU time.

> However, I was wandering, if anybody has any suggestions, what might
> be the performace decrease from 1 running job to two jobs by about 35%.
> Does that have to do with the 64GB kernel, since I do not notice this
> decrease on an Athlon MP Dual box, but using an Linux 2.4.18-4GB-SMP
> kernel?

Are the threads in fact being dispatched simultaneously on two CPUs, and is
this SMP machine really tuned for compute-intensive applications?

It sounds to me like the two threads are not being dispatched on both CPUs
but are contending for just one.

 
 
 

SMP scalabilty

Post by B. Joshua Rose » Thu, 26 Jun 2003 01:55:05



> We installed a vanilla Suse 8.1 on  Dell Xeon 2.4Ghz Dual box
> with 1GB Ram. The running kernel is a Linux 2.4.19-64GB-SMP #1 SMP.
> Now I tested the system with rather computational intensive numerical
> simmulations. If I run a single job, it takes about 1h 20min to finish.
> However, when I launching two identical jobs, these jobs need now 2h!
> Alternativly, I can launch the same job but runing the calculation in
> two threads, using pthread. Then, the job is finished in about 1hour,
> indicating a performance of ca 195% CPU time.

> However, I was wandering, if anybody has any suggestions, what might
> be the performace decrease from 1 running job to two jobs by about 35%.
> Does that have to do with the 64GB kernel, since I do not notice this
> decrease on an Athlon MP Dual box, but using an Linux 2.4.18-4GB-SMP kernel?

> Any suggestings appreciated.

>   Frithjof

If you only have 1G on that machine you should be using the 4G SMP kernel
not the 64G version. There is a performance penalty for using 64G address
translation. I doubt this has anything to do with your SMP performance
problems. One thing that might be causing the problem is Hyperthreading.
Hyperthreading causes a 25% decrease in single job performance, it's only
valuable if you are running lots of simultaneous threads. If you turn off
hyperthreading in your BIOS you will see a big jump in performance on your
two job test.
 
 
 

SMP scalabilty

Post by Bill Unr » Thu, 26 Jun 2003 02:14:18


]We installed a vanilla Suse 8.1 on  Dell Xeon 2.4Ghz Dual box
]with 1GB Ram. The running kernel is a Linux 2.4.19-64GB-SMP #1 SMP.
]Now I tested the system with rather computational intensive numerical
]simmulations. If I run a single job, it takes about 1h 20min to finish.
]However, when I launching two identical jobs, these jobs need now 2h!
]Alternativly, I can launch the same job but runing the calculation in
]two threads, using pthread. Then, the job is finished in about 1hour,
]indicating a performance of ca 195% CPU time.

Depends on whether or not the job is memory or cpu limited. Sounds like
memory. An smp machine cannot do anything about memory bandwidth and one
job can sit there waiting for the other cpu to finish.

]However, I was wandering, if anybody has any suggestions, what might
]be the performace decrease from 1 running job to two jobs by about 35%.
]Does that have to do with the 64GB kernel, since I do not notice this
]decrease on an Athlon MP Dual box, but using an Linux 2.4.18-4GB-SMP kernel?

]Any suggestings appreciated.

]  Frithjof

 
 
 

1. Scalabilty problem fixed width font

I use KDE 2.2 and a Courier New truetype font for fixed width text.

I have enabled anti-aliasining.

In Konqueror when a webpage has fixed width fonts embedded, some of the
letters aren't rendered completeley. Especially the O and m are incomplete
on the right side of the symbol.

I have made a test document in KWord containing text in Courier New in
various sizes. Most of the sizes don't display the above mentionend
problem. However at 10 pt. the problem exists. So it's a system wide
problem and not just Konqueror.

Does anyone else have this problem? Is there something I can do to solve it?

Bob

2. Squid Proxy Server ?

3. SMP - zombie processes with SMP enabled ...

4. wuftpd hangs sometimes

5. SMP motherboards (760 MPX chipset) and SMP howto

6. Access for people with disabilites:Mouse

7. Booting SMP kernel on non-SMP machines?

8. 3M DC 300A 2.9MB 300ft

9. Best NIC card -- LInux 2.4 SMP -> LInux 2.4 SMP

10. assertion failure : ext3 & lvm , 2.4.17 smp & 2.4.18-ac1 smp

11. FreeBSD 3.1 SMP outperforms SuSE Linux 6.0 SMP by 230 percent

12. SMP: Tyan Tomcat III dual, only one CPU active in /proc/smp

13. Pentium Pro SMP - scsi disks timing out at boot, only with SMP