Re: Pentium M and Oracle Performance

Hallo Andy,

[EMAIL PROTECTED] (none) writes:
> I hypothesize that there are two factors influencing the performance you
> see, and I do wish you had a better micro-view of the application so
> that we could get some real answers.

I can get into application details. However, when profiling the
application, it spends about 99% of its time in the oracle OCI
functions. I could provide parts of an Oracle trace file, the
Oracle initialization parameters, etc.

> Let me recap the data:  Your application runs on top of Oracle, and you
> have a standard benchmark suite which you have run on a variety of
> hardware including
>  - UltraSPARC-based UE x000-series systems

No. At customer sites, there are a few Siemens/Fujitsu systems,
though. Our Sparc/Solaris development system is just a Ultra 5/400
with an IDE disk. At least it has 2MB L2 cache.

>  - Xeon SMP systems with RAID
>  - Pentium M 1.6GHz laptop, 1MB cache, 2.5" IDE drive
>  - Pentium M 1.6GHz laptop, 1MB cache, 3.5" external USB2 drive

Let me list a few results. "Import" the the elapsed import time
in seconds for the Benchmark database. This has a closer correlation
to disk (write) performance than our code, and should not be too
dependant on out data. All systems ran Oracle 8.1.7.4.0.

Score Import System
 133    70   F/S Lifebook E4010, Centrino 1.6GHz, 512MB RAM, external 200GB USB2 disk, 
Win XP
 116    90   F/S Lifebook E4010, Centrino 1.6GHz, 512MB RAM, 60 GB 2.5" IDE, Win XP
 114    47   Compaq ProLiant ML530 G2, 2xPentium4 Xeon 2.4GHz, 3GB RAM, 10x36GB RAID1, 
Win2K
 109    56   Dell Dimension 4600, Pentium4 2.6GHz, 512MB RAM, IDE disk, Win XP
  87    77   F/S Primergy, 2xPentium3 1266MHz, 2GB RAM, EMC^2 storage array, Win2K
  82    52   Compaq ProLiant ML370 G3, 2xPentium4 Xeon 2GHz, 1GB RAM, HP Smartarray 
5302/128 (6x36GB RAID 5), Win2K
  45   135   HP rp5400, PA8600 540MHz, 1GB RAM, ?SCSI (no RAID), HP-UX 11
  33   127   Sun Fire 280R, 2xUSIII 750MHz, 2GB RAM, SUN StoreEdge A1000, Solaris 8
  21   237   Sun Ultra5/400,  USIIi 400MHz, 256MB RAM, 9GB IDE, Solaris 8
  20   242   HP Brio, PentiumII 350MHz, 256MB RAM, 2x9GB SCSI, Linux

I've omitted systems where the disk system was obviously dysfunctional
(including our only Sparc64GP result).

> The second-order determiner is cache size.  I theorize that your
> application has a sweet spot somewhere under 1MB, where the *vastly*

Is I said above, that would mean that Oracle (as we configure it)
has a sweet spot there. Our application code is lot in the noise.
As it happens, we like to set the Oracle log_buffer to 1MB for
the benchmark.

> It would be very interesting for you to run the benchmark under windows
> using the Intel VTune system to measure L2 cache misses, Icache
> activity, and I/O activity.

That would be some project... Would oprofile under Linux do the
trick, too? I have some chance of gettig that to work. Buying
VTune just to satisfy my curiosity is no option. After all,
we have no trouble with those results - just the customers
buying expensive disk systems should be somewhat annoyed.

Regards,
Chris