Anyone make 8 way Opteron Server?

[alent1234]

My company is looking to upgrade our 8 way Xeon boxes with dual opteron boxes. HP maxes out at a 4 way Opteron, but on AMD's website they say that it's possible to scale to an 8 way box. Anyone know of anyone that makes these. We are going to run windows 2003 64 bit on them along with SQL 2005.

 

[Hardlin]

Do you mean something like this:
Tyan
Or the Iwill version

 

[alent1234]

we were hoping for a rack mountable unit, but something like this. Is there an OEM like HP that makes these so we can buy one with support, etc?

 

[Hardlin]

I just gave you the original manufacturer so you could contact them to see who uses their base for servers. If you plug this into Google you will get some vendors but I haven't used them: 8 way opteron server

Also, these platforms support Dual Core so you could get 16 cores in one of them.

 

[Hardlin]

Here is a list of all the certified server partners from AMD:

http://www.amd.com/us-en/Processors/Pro...formation/0,,30_118_8796_10080,00.html

 

[Accord99]

Opterons don't scale too well past 4 sockets in typical server applications without additional logic, a well executed 4 socket solution like HP's shouldn't be much slower, while the 8 socket Xeon IBM x460 will be much faster.

 

[Markfw]

Originally posted by: Accord99
Opterons don't scale too well past 4 sockets in typical server applications without additional logic, a well executed 4 socket solution like HP's shouldn't be much slower, while the 8 socket Xeon IBM x460 will be much faster.

Do you have any proof of that ? The reviews I have read (no I don;t have links, just memory) said that the more cpu's, the more the Opterons walked away from the Xenon's due the the integrated memory controller. And I have seen 8-way 4U Opteron servers, that could now be 16 core.

 

[Topweasel]

Originally posted by: Markfw900

Originally posted by: Accord99
Opterons don't scale too well past 4 sockets in typical server applications without additional logic, a well executed 4 socket solution like HP's shouldn't be much slower, while the 8 socket Xeon IBM x460 will be much faster.

Do you have any proof of that ? The reviews I have read (no I don;t have links, just memory) said that the more cpu's, the more the Opterons walked away from the Xenon's due the the integrated memory controller. And I have seen 8-way 4U Opteron servers, that could now be 16 core.

Not Just the integrated Memory controller but the Point to Point bus, and NEMA all would separate a Opteron from a Xeon as you add cores. an 8 Way Dual core Opteron would kill a 8-or 16way Xeon setup no question.

 

[stardrek]

Here you are: http://www.sun.com/servers/entry/v40z/index.jsp

I have a few of these where I work, infact I just set one up today. They work very well. You can enjoy 4 dual core opterons in these puppies. They don't touch a Sun Fire 15k or 25K but they are still pretty sweet for the things you wanna do.

And YES, they will run Windows server 2003.

 

[MerlinRML]

Originally posted by: stardrek
Here you are: http://www.sun.com/servers/entry/v40z/index.jsp

I have a few of these where I work, infact I just set one up today. They work very well. You can enjoy 4 dual core opterons in these puppies. They don't touch a Sun Fire 15k or 25K but they are still pretty sweet for the things you wanna do.

And YES, they will run Windows server 2003.

Totally agreed. I like these boxes a lot, but they don't have any cable management which makes them less fun to work on.

I've been looking for a good AMD solution out there. I've looked at HP, IBM, and Sun. I haven't found anything that meets what I'm looking for, so I'm still looking, which is essentially all the goodness that an Intel-based server has but with an AMD-based solution. It's getting better, but it's just not quite there yet.

I have not seen an 8-way solution by any of the big vendors. The closest I've seen is a quad dual-core.

 

[Rubycon]

I know someone that has one and you need earplugs to be near the thing. :Q

 

[alent1234]

hopefully we can get a test HP 4 way Opteron to do some stress testing in the near future. We'll probably hold out for HP or buy the 4 way box. Looked at the Sun boxes, but didn't see anything about official support for windows 2003. Didn't find anything good with the others.

 

[stardrek]

Originally posted by: alent1234
Looked at the Sun boxes, but didn't see anything about official support for windows 2003.

The box i suggest is officially supported by MS: http://www.microsoft.com/windows/catalo...chType=gs&qu=V40z&scope=1&btnSearch=Go

It also says it on the link I gave. Just thought you should know.

 

[imported_Tick]

What is it your running that you need that much processing density?

 

[alent1234]

a few instances of SQL

right now our main SQL boxes are 8 way Xeon 900MHz boxes with 2MB cache on the chip. THey were the rage 3 years ago. Business is expanding this year and they are being taxed pretty heavy. Need faster boxes to run SQL. We are thinking 4 - 8 way Opteron's and 16GB - 32GB RAM.

 

[Accord99]

Originally posted by: Markfw900

Originally posted by: Accord99
Opterons don't scale too well past 4 sockets in typical server applications without additional logic, a well executed 4 socket solution like HP's shouldn't be much slower, while the 8 socket Xeon IBM x460 will be much faster.

Do you have any proof of that ? The reviews I have read (no I don;t have links, just memory) said that the more cpu's, the more the Opterons walked away from the Xenon's due the the integrated memory controller. And I have seen 8-way 4U Opteron servers, that could now be 16 core.

It's the nature of the architecture. For every memory access, an Opteron cpu has to ask every other cpu in the system to see if there's a modified version in its cache. In an optimal 8 socket system, the farthest cpu is still 3 hops away, each hop adding latency. In most of the current 8 socket systems, like the Iwill and Tyan they're not optimal and you end up with 4 hops. IBM uses more advanced cache coherency that allows it to scale to 32 sockets. It powers the fastest x86 server in TPC-C, the most important database benchmark.

 

[Matthias99]

Originally posted by: Accord99

Originally posted by: Markfw900

Originally posted by: Accord99
Opterons don't scale too well past 4 sockets in typical server applications without additional logic, a well executed 4 socket solution like HP's shouldn't be much slower, while the 8 socket Xeon IBM x460 will be much faster.

Do you have any proof of that ? The reviews I have read (no I don;t have links, just memory) said that the more cpu's, the more the Opterons walked away from the Xenon's due the the integrated memory controller. And I have seen 8-way 4U Opteron servers, that could now be 16 core.

It's the nature of the architecture. For every memory access, an Opteron cpu has to ask every other cpu in the system to see if there's a modified version in its cache. In an optimal 8 socket system, the farthest cpu is still 3 hops away, each hop adding latency. In most of the current 8 socket systems, like the Iwill and Tyan they're not optimal and you end up with 4 hops. IBM uses more advanced cache coherency that allows it to scale to 32 sockets. It powers the fastest x86 server in TPC-C, the most important database benchmark.

Do you have a list of 'official' TPC-C x86 results? The only ones I can find are the 'overall' ones, and those are indeed led by IBM -- using 64-way POWER5 processors. Totally different architecture, and not really comparable.

Here are two sample results out of the TPC-C database:

IBM - 8-way Xeon MP 3.3Ghz
HP - 4-way Opteron 2.4Ghz

The 4-way Opteron server is only a hair slower, but costs a third as much. Adding four more CPUs to the Opteron server would result in it just blowing away the Xeons.

The problem with multiway Xeon servers is that, even though the Opterons may have slightly higher latency at the number of CPUs scale up, the Xeons are throttled in terms of bandwidth. For anything heavy on memory bandwidth, they just won't perform well. The latest ones try to remedy this somewhat by upping the FSB speed, but I'm unconvinced that it would really be enough.

 

[Accord99]

Originally posted by: Matthias99
Do you have a list of 'official' TPC-C x86 results? The only ones I can find are the 'overall' ones, and those are indeed led by IBM -- using 64-way POWER5 processors. Totally different architecture, and not really comparable.

Here are two sample results out of the TPC-C database:

IBM - 8-way Xeon MP 3.3Ghz
HP - 4-way Opteron 2.4Ghz

The 4-way Opteron server is only a hair slower, but costs a third as much. Adding four more CPUs to the Opteron server would result in it just blowing away the Xeons.

That's a 4 Socket, dual-core Opteron system. That system is also unable to add 4 more CPUs, unlike the IBM Xeon system. Nearly all of the cost difference comes from the storage portion.

The problem with multiway Xeon servers is that, even though the Opterons may have slightly higher latency at the number of CPUs scale up, the Xeons are throttled in terms of bandwidth. For anything heavy on memory bandwidth, they just won't perform well. The latest ones try to remedy this somewhat by upping the FSB speed, but I'm unconvinced that it would really be enough.

The IBM X3 chipset has 4 dual-channel memory controllers and dual FSBs per 4 socket quad and has 2 6+GB ports to connect with other X3 quads.