Dual Processor Pentium 4 runs Slow!

[brodin]

When I run a single process of Prime95 it has a CPU usage of 50% and a Per iteration time: ~0.045 sec.

When I run 2 processes (one on CPU 0 and the other on CPU 1) I see a CPU usage of 100% (50% for each process) but each process has a Per iteration time: ~0.102 sec. This is slower than a single process.

As a test I ran 2 processes on a single processor. This shows a CPU usage of 50% (25% for each process) as expected. The Per iteration time for each process is ~0.090 sec (as expected).

Why do 2 processes (1 on each of 2 processors) run slower than 2 processes on the same processor, and even slower than a single process ???

(I know which processors are being used by looking at the graphs of the "CPU Usage History" in Windows Task Manager.)

This is a dual processor Pentium 4 under Windows XP.

How can I get the proper throughput for 2 processors ?

[R.D. Silverman]

Originally Posted by brodin

When I run a single process of Prime95 it has a CPU usage of 50% and a Per iteration time: ~0.045 sec.

When I run 2 processes (one on CPU 0 and the other on CPU 1) I see a CPU usage of 100% (50% for each process) but each process has a Per iteration time: ~0.102 sec. This is slower than a single process.

As a test I ran 2 processes on a single processor. This shows a CPU usage of 50% (25% for each process) as expected. The Per iteration time for each process is ~0.090 sec (as expected).

Why do 2 processes (1 on each of 2 processors) run slower than 2 processes on the same processor, and even slower than a single process ???

(I know which processors are being used by looking at the graphs of the "CPU Usage History" in Windows Task Manager.)

This is a dual processor Pentium 4 under Windows XP.

How can I get the proper throughput for 2 processors ?

The short answer is: "Bus Contention"; both processes are competing to
move data from main memory to the L2 cache.

Note that when you have two processes running even though each takes
.102 sec/iteration you are producing twice the output, so the effective
throughput is .051sec/iteration; about 20% slower than just one process.

If you think two copies of Prime95 has problems try running two copies
of NFS

[patrik]

Originally Posted by brodin

This is a dual processor Pentium 4 under Windows XP.

Are there dual processors for Pentium 4? Or do you mean a Xeon? I just want to make sure you are not confusing it with one hyperthreaded processor, which (I've heard) shows as two logical processors.

[brodin]

Originally Posted by patrik

Are there dual processors for Pentium 4? Or do you mean a Xeon? I just want to make sure you are not confusing it with one hyperthreaded processor, which (I've heard) shows as two logical processors.

I now think that I really only have one processor in my Pentium 4, but it shows up in the Windows Task Manager as two processors. How can I tell which I really have? Also, why would they have it show up as two processors, when it is really only one? Thanks.

[PrimeCruncher]

Originally Posted by brodin

Also, why would they have it show up as two processors, when it is really only one? Thanks.

HyperThreading creates a second 'logical' processor which most apps can use like a normal physical processor. However, it doesn't create any new floating point registers or pipelines, so Prime95 can't really take advantage of it. There is apparently some gain to be had by running LL tests on the 'physical' processor and TFs on the 'logical' one, but it's tricky to implement. There are numbers on this in various other threads if you want to look at them...

[Moloch]

here are a couple of times for LL/TF trials I have run on my own p4 machine.

Times for LL/TF running a single application are:

.078 sec for LL
.0096 sec for TF at 63 bits
.0117 sec for TF at 66 bits

Times running 2 apps - LL on p0 and TF on p1:

.137 sec for LL (TF at 63 bits) - 14% gain
.0159 sec for TF at 63 bits - 21% gain
*overall gain at 63 bits - 35%

.141 sec for LL (TF at 66 bits) - 10.5% gain
.0165 sec for TF at 66 bits - 42% gain
*overall gain at 66 bits - 52.5%

Machine stats:
P4 3.0C @ 3.3GHz
FSB @ 220 Mhz (880 quad pumped)
MoBo - ASUS P4C800-E
RAM - 1GB Corsair Twin-x 4400 (512x2)
Raid0 - 2 x 80GB Seagate Barracuda


If you are trying to do 2xLL tests I dont think you will see any improvement (probably a slowdown of both). Running a LL/TF combination shows substantial gains since they use different parts of the processor. Each app will take a little longer to complete, LL will take nearly twice as long while TF only about 50% longer, but if you plan on running both types of tests anyway this will save time in the long run.

If you are interested in running 2xTF apps at once make sure one is set to 64 bits and below and the other to 65 and above. I havent tested TF/TF myself, but from what I hear that is the only way to show improvement.