Advanced Micro Devices Inc (AMD)

[mva5]

WBMW - This is what Dan must have meant when he said "Ha Ha Ha":

Comments on SPECmanship courtesy of Eachus (fool.com msg):

But here is a sample from what is available, three 4 CPU results for XeonMP 2.8 GHz, Opteron 846, and Itanium2 1.5 GHz:

Dell PowerEdge 6650 4x2.8 GHz XeonMP, Zeus 4.2r2, SPECweb99_SSL = 2177

AMD Appro 4144H 2.0 GHz Opteron, Zeus 4.2r2, SPECweb99_SSL = 3399

HP Integrity rx5670 4x1.5 GHz Itanium2 6M, Zeus 4.2r2, SPECweb99_SSL = 3702

Looks like the Opteron does significantly better than the Xeon, but not as well as the Itanium2. Or does it? The devil is always in the details, and I am going to choose just two:

Dell Xeon:Memory 16 GB, Disk Subsystem 1 36GB, 4-18GB 15KRPM drives

AMD Opteron:Memory 32 GB, Disk Subsystem 3-18GB 15KRPM Ultra 320 SCSI DrivesHP

Itanium2:Memory 32 GB, Disk Subsystem 3x73GB (15K RPM) 1x36GB (15K RPM) 15x18GB(15K RPM) HP VA7100 disk array

The XeonMPs have to cope with PAE overhead, but apparantly on this test it is worth it to have the "extra" memory. The Opteron test system used PC2100 memory instead of PC2700, but I think we would find that is the same issue as with the Xeon--you could use PC2700 memory but less of it, and more main memory is more important for this benchmark than fast main memory. This is not surprising since this is a database benchmark and caching data from disk is going to have a high payoff.

But look at the disk system for the Itanium2! The AMD system uses only 54 GBytes of disk, and the Xeon system twice that, so why does the Itanium2 system have so much disk storage, 525 Gigabytes, almost ten times as much as the Opteron system? Since the SPEC requirements for documentation are pretty good, we can find:

HP:
(73 GB) disk used for OS
(73 GB) disk used for doc_root
(73 GB) disk used for the file_set which is mounted read-only
(36 GB) disk with largefiles, behind used for post.log
HP VA7100 disk array with dual controllers (A6188A) with high performance enabled for web log# web log file system is striped using 64K stripes over 10 LUNs and mounted with largefiles and behind options

Work it all out, and what that really says is that the web log is spread over ten disks. Since during the test the web log is effectively a write only file, this probably means that the web log writes are always to the (probably 2 Meg) caches that are part of the individual disks.

Would you set a system up like that? Not me. In fact notice that the html files are mounted read only! It would be much cheaper to buy more CPUs. Is this normal SPECmanship or carried to extremes? That is up to you to judge. I just note it as part of the liberal dose of salt I take with these tests.

There are some SPECmanship games everyone is playing. Dell either didn't know about the trick of mounting the html files as read only or it didn't help on their system. (Since they put the fileset and logs on the same RAID0 volume, it may be that the RAID controller they used didn't support multiple volumes per disk--or that may have resulted in a lower score.

Link for Spec results:

http://www.spec.org/web99ssl/results/web99ssl.html

Original post by Eachus:

http://boards.fool.com/Message.asp?mid=19916881