8 way processor vs. 4 way processor in regards to SQL

[cfadool]

Hello. Got a question for everyone here.

My company has asked me to research the performaace differences between and 8 way processor vs. 4 way processor server.

The application housed on this server will be an application that relies heavily on SQL 2000. The cost of an 8 way proc. server is a huge, huge difference in regards to a 4 way proc.

Microsoft (as MS does) is telling those in charge that it will make a huge difference. However, they cannot seem to provide any hard data to support this.

Does anyone have any experience with an 8 way processor server AND can provide any sort of whitepaper or a link to one?

Thanks in advance!

 

[Brucmack]

The most likely beneficial information I could come up with are benchmarks IBM has done for its servers. The main entry point is here:

http://www.pc.ibm.com/ww/eserver/xseries/benchmarks/series.html

Most of the database benchmarks are DB2, but they have some SQL Server 2000 benchmarks for 8-way and 4-way servers. They aren't necessarily at the same clock speed though :/

A sample of what you can find:

8-way Xeon 3.0 GHz: 156,105.72 tpmC
4-way Xeon 2.8 GHz: 90,271.76 tpmC

However, you'll have to do some more digging to see if there are any other architectural improvements there that account for the increase. In any case, it's about a 70% increase for about double the cost.

 

[cfadool]

 

[Matthias99]

AT actually put up a review a few days ago about 2-way versus 4-way database servers, and made some extrapolations about how they would scale to 8-way. Not quite what you're looking for, but it may help. Their results were essentially that you want Opterons, not Xeons, for multiway database servers (at least until 800Mhz Xeon MPs come out, and even then the Opteron 848s will still probably cost half as much).

AT Review

Unless you absolutely NEED to have a single server with more horsepower than a 4-way can provide, it's a game of diminishing returns (that is, it costs more than twice as much for less than twice as much performance each time you double the number of CPUs). You'd generally be far better off (financially, at least) with two 4-way servers working off a common dataset than a single 8-way one. Don't know how feasable that is for your setup, though.

 

[glugglug]

The scaling from 4 to 8 way on Intel chips will NOT be what you expect, and will greatly depend on the manufacturer of your particular 8 way box.

In a 4-way system pretty much every manufacturer will take the path of least resistance and use the standard chipsets and reference designs available, in which the 4 CPUs are sharing a system bus.

No standard Intel chipset supports more than 4 CPUs. In fact the only x86 compatible chip _ever_ really designed to be put in an 8-way configuration without a lot of extra work for the mobo designer is the Opteron. This means every 8-way P4 server out there is a custom design, not a reference board. (Hence the much higher price) This generally means at least double the bus bandwidth (as in a design where the 8-way system is essentially a pair of 4-way systems with some kind of bridge in-between them), sometimes a lot more (full NUMA architecture). So the difference between a 4-way and 8-way system is actually proportionally greater than the difference between a 2-way and 4-way system.

 

[cfadool]

Thanks for some great information. That is what I thought and was telling them.

I appreciate the information.

 

[Ghostt]

also remember that the fastest opterons cost less then half of the fastest xeons

from anand's article

The 533MHz FSB 2MB L3 Prestonia based Xeon manages to help Intel tremendously in keeping competitive with the Opteron. In fact, under heavy enough workloads there is virtually no performance difference between a 3.2GHz Xeon and a 2.2GHz Opteron (x48). It isn't until you move to 4-way configurations that AMD's platform architecture begins to flex its muscle. That being said, Intel has done an incredible job of keeping up performance wise in 2-way configurations; we have a much better showing here than we did in the web server test.

Interestingly enough, while the new Gallatin Xeon MPs have a massive 4MB L3 cache, most of that cache will end up being used to keep traffic off of the bandwidth starved 400MHz FSB. The performance gap between the Opteron 848 and the Xeon MP is amplified significantly once you move to a 4-way setup; the Xeon's shared bus just can't cut it anymore, not at 400MHz. AMD's point-to-point Hyper Transport implementation helps extend their performance advantage significantly. An 8-way Opteron vs. Xeon comparison would not be pretty.

The comparison we've made here is a very important one; it identifies Intel's strengths and their weaknesses with Xeon, and it crowns Opteron a clear multiprocessor winner. An area that we didn't touch on is cost, which is where AMD truly shines. The Opteron 848 processors we tested are around 1/2 the price of Intel's 2MB L3 Xeon MPs and we have not seen retail data on how expensive the 4MB parts will be.

In a 4-way configuration AMD's Opteron cannot be beat, and thus it is our choice for the basis for our new Forums database server. We'll be documenting that upgrade in a separate article so stay tuned.

read the whole article for the real story

AMD Opteron vs. Intel Xeon: Database Performance Shootout

 

[cfadool]

 

[SuperTool]

Get the 4 way. Why? Because it's Intel. There is a reason they only support 4 way.
Do you think if their SMP systems actually scaled well beyond 4 way they would not develop a chip for that?
This is a company that spent a $1B+ on Itanic, probably in vain, to compete in the server market. Do you think they wouldn't spend a few mil designing an ASIC to support more than 4 way SMP if there was actual sense in doing so?

 

[Sohcan]

Originally posted by: SuperTool
Get the 4 way. Why? Because it's Intel. There is a reason they only support 4 way.
Do you think if their SMP systems actually scaled well beyond 4 way they would not develop a chip for that?

Really? The enterprise benchmarks tell otherwise...let's look at TPC-C. Results using 1.5 GHz Itanium 2:

Unisys ES7000 Aries 420 Enterprise Server (16-way): 309036.53 TpmC
NEC Express5800/1320Xd (32-way): 577530.77 TpmC (+87% over 16-way)
HP Integrity Superdome (64-way): 1008144.49 TpmC (+75% over 32-way)
...or, if you want to use a clustered solution:
HP Integrity rx5670 Cluster 64P: 1184893.38 TpmC (+105% over best 32-way)


How about SAP SD 2-tier?

HP Integrity rx4640 (4-way): 715
HP Integrity rx8620 (16-way): 2880
Whoops, that's superlineal scaling.


SPECjbb2000:

HP Integrity Superdome (16-way): 322604
HP Integrity Superdome (32-way): 574912 (+78%)
HP Integrity Superdome (64-way): 1008604 (+75%)

As a comparison, the 1.35 GHz SPARC64 V:
Fujitsu PRIMEPOWER1500 (32-way): 492683
Fujitsu PRIMEPOWER2500 (64-way): 835479 (+70%)

Not an enterprise benchmark, but here's SPEC CPU rate:
SPECint_rate:
SGI Altix 3000 (8-way): 98.3
SGI Altix 3000 (16-way): 195 (+98.4%)
SGI Altix 3000 (32-way): 385 (+97.4%)
SGI Altix 3000 (64-way): 854 (+122%, there's that superlineal scaling again)

SPECfp_rate:
SGI Altix 3000 (4-way): 82.2
SGI Altix 3000 (8-way): 164 (+97.1%)
SGI Altix 3000 (16-way): 327 (+99.4%)
SGI Altix 3000 (32-way): 644 (+96.9%)
SGI Altix 3000 (64-way): 1250 (+94.1%)

Do you think they wouldn't spend a few mil designing an ASIC to support more than 4 way SMP if there was actual sense in doing so?

Intel's E8870 chipset supports up to 8 Itanium 2s.

 

[SuperTool]

Originally posted by: Sohcan
<blockquote>Quote
Originally posted by: SuperTool
Get the 4 way. Why? Because it's Intel. There is a reason they only support 4 way.
Do you think if their SMP systems actually scaled well beyond 4 way they would not develop a chip for that?

Really? The enterprise benchmarks tell otherwise...let's look at TPC-C. Results using 1.5 GHz Itanium 2:

Unisys ES7000 Aries 420 Enterprise Server (16-way): 309036.53 TpmC
NEC Express5800/1320Xd (32-way): 577530.77 TpmC (+87% over 16-way)
HP Integrity Superdome (64-way): 1008144.49 TpmC (+75% over 32-way)
...or, if you want to use a clustered solution:
HP Integrity rx5670 Cluster 64P: 1184893.38 TpmC (+105% over best 32-way)


How about SAP SD 2-tier?

HP Integrity rx4640 (4-way): 715
HP Integrity rx8620 (16-way): 2880
Whoops, that's superlineal scaling.


SPECjbb2000:

HP Integrity Superdome (16-way): 322604
HP Integrity Superdome (32-way): 574912 (+78%)
HP Integrity Superdome (64-way): 1008604 (+75%)

As a comparison, the 1.35 GHz SPARC64 V:
Fujitsu PRIMEPOWER1500 (32-way): 492683
Fujitsu PRIMEPOWER2500 (64-way): 835479 (+70%)

Not an enterprise benchmark, but here's SPEC CPU rate:
SPECint_rate:
SGI Altix 3000 (8-way): 98.3
SGI Altix 3000 (16-way): 195 (+98.4%)
SGI Altix 3000 (32-way): 385 (+97.4%)
SGI Altix 3000 (64-way): 854 (+122%, there's that superlineal scaling again)

SPECfp_rate:
SGI Altix 3000 (4-way): 82.2
SGI Altix 3000 (8-way): 164 (+97.1%)
SGI Altix 3000 (16-way): 327 (+99.4%)
SGI Altix 3000 (32-way): 644 (+96.9%)
SGI Altix 3000 (64-way): 1250 (+94.1%)

<blockquote>Quote
Do you think they wouldn't spend a few mil designing an ASIC to support more than 4 way SMP if there was actual sense in doing so?[/quote]Intel's <a target=new class=ftalternatingbarlinklarge href="http://www.intel.com/design/chipsets/e8870sp/index.htm?iid=ipp_srvr+chpsts_e8870sp&amp;">E8870 chipset</a> supports up to 8 Itanium 2s.[/quote]

I thought he was asking about Xeons, not Itanics.