RS/6000 vs. SP2 performance

RS/6000 vs. SP2 performance

Post by Pascal Sartorett » Fri, 15 Oct 1999 04:00:00



Hello,

Our company has developed an AIX application which does image
processing for satellite data. It uses approx. 512 MB of memory to run.
It has been developed on an RS/6000 workstation (a 43P), and shall
then run on a SP2 node after. We had no problem in porting RS/6000 =>
SP2, except for performance:

The published benchmarks (SPECInt and the like) of the 2 machines
roughly say that an SP2 node is twice as fast for real computations
and twice as slow for integer computations. Our application uses both,
maybe more integer numbers, hence we expected as worse case a twofold
loss in performance on the SP2.

The actual results are very different: the SP2 is 5-6 times slower
as the RS/6000, with a similar memory configuration.

Does anybody have had a similar experience? Other benchmarks, or
even a rule of thumb?

One possible explanation we heard: the SP2 has no level-2 cache. Any
other idea?

Thanks in advance for even the slightest information....

Pascal Sartoretti

 
 
 

RS/6000 vs. SP2 performance

Post by Pascal Sartorett » Fri, 15 Oct 1999 04:00:00


Hello,

Our company has developed an AIX application which does image
processing for satellite data. It uses approx. 512 MB of memory to run.
It has been developed on an RS/6000 workstation (a 43P), and shall
then run on a SP2 node after. We had no problem in porting RS/6000 =>
SP2, except for performance:

The published benchmarks (SPECInt and the like) of the 2 machines
roughly say that an SP2 node is twice as fast for real computations
and twice as slow for integer computations. Our application uses both,
maybe more integer numbers, hence we expected as worse case a twofold
loss in performance on the SP2.

The actual results are very different: the SP2 is 5-6 times slower
as the RS/6000, with a similar memory configuration.

Does anybody have had a similar experience? Other benchmarks, or
even a rule of thumb?

One possible explanation we heard: the SP2 has no level-2 cache. Any
other idea?

Thanks in advance for even the slightest information....

Pascal Sartoretti

 
 
 

RS/6000 vs. SP2 performance

Post by milline » Fri, 15 Oct 1999 04:00:00



Quote:> Hello,

> Our company has developed an AIX application which does image
> processing for satellite data. It uses approx. 512 MB of memory to run.
> It has been developed on an RS/6000 workstation (a 43P), and shall
> then run on a SP2 node after. We had no problem in porting RS/6000 =>
> SP2, except for performance:

> The published benchmarks (SPECInt and the like) of the 2 machines
> roughly say that an SP2 node is twice as fast for real computations
> and twice as slow for integer computations. Our application uses both,
> maybe more integer numbers, hence we expected as worse case a twofold
> loss in performance on the SP2.

> The actual results are very different: the SP2 is 5-6 times slower
> as the RS/6000, with a similar memory configuration.

Things to check; 1) there are several flavours of SP node, both MCA and PCI
and uni- and multi-processor. Depending on which 43P you built on, you might
expect the SP node performance to be anywhere from worse than, comparable to
or better than the 43P.
2) Memory disposition in some nodes is significant due to memory bus design;
it's better to have 4 memory boards rather than 2 to get full memory
bandwidth.
3) SP nodes have more or less direct equivalents in standalone RS/6000
world, but there's no equivalent of 43P (H50 with 332Mhz processor is first
PCI node type).
4) If you haven't already, it might be worth recompiling your code in native
mode (i.e. processor specific). By default code builds in common mode, which
means that the executable will run on any RS/6000 processor (POWER, POWER 2,
POWER 3, POWER-PC 601/604/604e etc) without modification. This does mean
losing some optimisations specific to your CPU.
5) If your SP node is an SMP node, your application may not be
multi-threading, effectively tying you down to a single processor.

Regards,

Derek Milliner

This a personal account, don't blame my employers if I louse up
or just plain lie through my teeth...

 
 
 

RS/6000 vs. SP2 performance

Post by Peter Bjoe » Fri, 15 Oct 1999 04:00:00


On Thu, 14 Oct 1999 16:29:12 GMT a carbon-based biped by the name of Pascal


>The published benchmarks (SPECInt and the like) of the 2 machines
>roughly say that an SP2 node is twice as fast for real computations
>and twice as slow for integer computations. Our application uses both,
>maybe more integer numbers, hence we expected as worse case a twofold
>loss in performance on the SP2.

>The actual results are very different: the SP2 is 5-6 times slower
>as the RS/6000, with a similar memory configuration.

But there are many different kinds of SP nodes.
Some have a PowerPC chip, some have a Power2 or Power3
chip running at various clock speeds.
Which node type are we talking about here ?

Regards

Peter

 
 
 

1. rs/6000 Model 220 graphics performance (GT1 vs GT3)

Does anyone have a GT3 graphics adapter in a model 220 that could give me some
xperf results? We need to compare this with the GT1 adapter. The GT1 scrolling
speed seems to be very poor and makes the machine seem very slow.  In addition,
it seems to make the 220 go nearly 100% compute bound when scrolling alot of text
in a window.  Thanks in advance for any help...

-art stine
sr network engineer
clarkson u

2. linksys etherfast pci card does NOT work!

3. FA: RS/6000 SP2 System

4. SVGAlib and ThreedKit

5. Is CDE better over OpenWindow ? Thanks

6. hdd requirements for text mode

7. SP2, Rs/6000 Sr. Systems Administrator/Consultant

8. FA: RS/6000 SP2 System

9. Immediate job opp-RS/6000 SP2 adminstrator