Re: Performance issues

[Marcus G. Daniels]

Bill Northcott wrote:

> > Motherboards like Tyan Trinity GC-SL (for Pentium 4, not even a Xeon)
> > have PCI-X slots run 64 bits wide @ 133 Mhz. 
> >    
>
> They are not common.  The standard Intel chip sets only support 
> 32bit/33MHz PCI on P4 computers.
>  
If you want a server, buy a motherboard with a server chipset.
The performance of the Pentium 4 CPU _itself_ is a different issue.

> Latency is much more important in most applications.  Which is why high 
> end Macs and powerful number crunchers use Level 3 cache memory
Level 3 cache is a way to deal with the memory bus clock speed being 
limited and inherent RAM latency.  On a new and mainstream Pentium 
4-based system, the clock is four times as fast as on a Mac, so there is 
less benefit in a level 3 cache.  New Pentium 4s have 512KB on the chip 
itself, and they use very fast RAM, both in terms of bandwidth and latency.

For a benchmark, let me propose a garbage collecting test program, such 
as the Java version GCBench.  
http://www.hpl.hp.com/personal/Hans_Boehm/gc/gc_bench

Folks can compile the GCBench.java referenced here on different machines 
with minimum trouble and make some comparisons. 

I propose this benchmark because it will be traversing from pointer to 
pointer within a large memory region larger (say 32MB) than any 
potential Level 2 or 3 cache.  Information gets torn up and down in 
different ways, and should be a big stress on the caches.  That it is 
Java just makes it easy for people to run, but also has the benefit that 
the highly-tuned Sun JDK generational garbage collector has the 
opportunity to sort objects (and thus memory acceses) by age and 
physical region, possibly leveraging the available cache.  

Results:

On a 550 Mhz Pentium 3 with 256K Level 2 cache and PC100 memory I get 
the following:

time jdk1.3.1_07/bin/java -Xms32m -Xmx32m GCBench
[output deleted]
real    0m8.565s
user    0m8.280s
sys     0m0.310s

This is roughly comparable to a Sun Blade 100 @ 500 Mhz also with 256K 
Level 2 cache.

time /opt/jdk1.3.1/bin/java GCBench
[output deleted]
real    0m10.609s
user    0m9.330s
sys     0m0.680s  

On a 400 Mhz G4 with 1MB Level 2 cache and PC100 memory I get 6.7 user 
seconds of  the built-in JDK 1.3.1 with Mac OS X 10.2.3.

time java -Xms32m -Xmx32m GCBench 
[output deleted]
real    0m9.183s
user    0m7.000s
sys     0m0.760s 

Very impressive.  The G4 is only running at 400 Mhz.  But then it has 
four times as much level 2 cache.  It _ought_ to.

Now how about a 1Ghz Pentium III with PC133 memory and 256K level 2 cache?
In other words, since the GCBench should be latency-limited to a 
considerable extent, the faster PC133 RAM should make a difference and 
the CPU won't help so much.  That would appear to be the case:

time /usr/java/jdk1.3.1_07/bin/java GCBench
[output deleted]
real    0m7.186s
user    0m6.970s
sys     0m0.090s

How does this compare to a Pentium 4 @ 1.7Ghz with fast Rambus memory 
and a 400 Mhz front side bus, and 256K level 2 cache?

real    0m2.799s
user    0m2.510s
sys     0m0.120s 

Still a small level 2 cache, but fast RAM and a fast CPU trounces the G4 
with four times the level two cache.

Since the current generation Pentium 4's have 512KB level 2 cache and 
DDR and Rambus is even faster now than with the test machine I used, it 
would be interesting to compare a 3.06 Ghz Pentium 4 with 512 KB level 2 
cache and PC2700 DDR memory (or 1066 RDRAM) against the latest Apple 
offering (Xserve, new Quicksilver, etc.).   (I don't have any of these 
at the moment.)   I predict that the Mac will be treated to a big dose 
of whoop-ass.

> and modern 
> RAID controllers boast Gigabytes of DRAM and very sophisticated caching 
> algoritms.
Well, $150 Pentium 4 motherboards with RAID 0 and two to four ports are 
common these days.  A 64 bit PCI (not even PCI-X) ATA RAID card will be 
a couple hundred dollars for a dual Pentium III motherboard.  If you 
want to spend more money, there are server motherboards for Pentium IVs 
that have the latest-and-greateast SCSI controllers.   Server 
motherboards seem to be four or five times as expensive as mainstream 
motherboards (e.g. $600).

Btw, another possible benchmark would be SQL server performance.  SQL 
servers use a lot of RAM and have access patterns that are unpredictable 
(thus stressing the data cache).  I tried to find some SQL numbers for 
the XServe, but of course Apple advocates rarely bother to publish 
benchmarks.  Reality distortion field and all that..


                 ==================================
  Swarm-Support is for discussion of the technical details of the day
  to day usage of Swarm.  For list administration needs (esp.
  [un]subscribing), please send a message to <address@hidden>
  with "help" in the body of the message.