Benchmarks MySQL: 64 bit versus 32 bit

Sixteen registers, more than 4GB physical and virtual memory, using 64 bit software should have nothing but advantages. To see how much advantage that the new 64 bit binaries offer, we tested both the Xeon Irwindale and the Opteron on 32 bit MySQL and 64 bit MySQL, both version 4.0.18.

The Opteron was tested on the MSI board for these tests, contrary to previous tests where the Iwill board was used. The Intel CPU was running on the Intel board as in all previous tests.

Concurrency Dual Xeon (Irwindale)
3.6GHz with HT 64 bit
Dual Xeon (Irwindale)
3.6GHz with HT 32 bit
Dual Opteron 248 64 bit Dual Opteron 248 32 bit Dual Xeon (Irwindale)
3.6GHz 64 bit versus 32 bit
Dual Opteron 248
64 bit vs 32 bit
1 286 245 324 261 16% 24%
2 450 379 532 421 19% 26%
5 497 534 642 485 -7% 32%
10 517 563 691 509 -8% 36%
20 545 631 692 527 -14% 31%
35 506 616 670 514 -18% 30%
50 495 559 666 516 -11% 29%
             
AVG 512 580 672 510 -12% 32%
MAX 545 631 692 527 -14% 31%

This is really remarkable, as the Xeon does not benefit from 64 bit at all. Worse, a 10% performance penalty is paid for moving over to 64 bit. The Opteron, however, thrives on 64 bit and gets a 30% boost from 64 bit.

Now, it is possible that the 64 bit binary is simply very well optimized for the Opteron. The 64 bit compiler used by the MySQL engineers (obviously not the Intel compiler, gcc) might not have the necessary optimisations to get the best out of the Xeon architecture. That is probably the most important reason why the difference (+30% versus - 12%) is so big.

However, when we take a look at the numbers in DB2, you will notice that the Xeon runs about 2 to 3% slower, while the Opteron gains a 12% boost from 64 bit. IBM's 32 bit binaries make the Xeon run as fast as the best Opterons. Once we turn to 64 bit binaries, the Opteron gets the upper-hand again. So, there is more: for some reason, the Xeon is not too happy with 64 bit binaries. We can only speculate, but maybe (some) 64 bit calculations have to cycle twice through the ALU's of the Prescott/Nocona/Irwindale architecture.

The consequence is that a Xeon running a 32 bit application is quite a bit faster than the competition, but once you switch to 64 bit, the Xeon does not stand a chance against the Opteron.

Benchmarks MySQL: Single core versus Dual core

Some of you might already get nervous: where is the dual core Opteron? SUSE SLES 9 Linux was a little more stubborn. With the original SLES 9 kernel 2.6.5-97, the dual Opteron would just crash. We applied Service Pack 1 (2.6.5-157smp) and the new Opteron would boot and recognize the two cores, but the second CPU was disabled because of APIC IRQ problems.

Therefore, we were only able to run the Dual core Opteron on Gentoo with a 2.6.12 kernel. The Iwill board still had trouble running two cores. We run the tests on the MSI board. To give you an idea of how Gentoo and the new kernel compare to SUSE SLES 9 SP1, and IWill K8ES to MSI's K8Master2-FAR, we ran a few tests with SUSE on the MSI board too.

Concurrency Dual Core Opteron 875 - MSI -
Gentoo
Dual Opteron 248 - MSI -
Gentoo
Dual Opteron 248
- MSI - SUSE
Dual Opteron 248 -
Iwill - SUSE
Dualcore vs Dual CPU Gentoo vs SUSE Iwill vs MSI
1 288 270 324 264 7% -17% -19%
2 463 443 532 461 4% -17% -13%
5 583 558 642 591 5% -13% -8%
10 616 601 691 670 2% -13% -3%
20 648 610 692 683 6% -12% -1%
35 664 611 670 659 9% -9% -2%
50 628 579 666 662 8% -13% -1%
AVG 628 592 672 653 6% -12% -3%
MAX 664 611 692 683 9% -12% -1%

SUSE SLES 9 SP1 is quite a bit faster than a standard tuned Gentoo installation. Some of the improvements in kernel 2.6.12 might have traded performance in for more stability.

The second CPU on the MSI board does not have its own local memory, and has to access the RAM via the Hypertransport connection to the crossbar switch of the first CPU. Just like one dual core Opteron, the two CPUs have to share the bandwidth of one dual channel memory bus. Therefore, the comparison of one dual core Opteron and two single Opterons at the same clock speed is very interesting: it gives us some insight on how much performance is gained by letting the two cores talk over the System Request Queue instead of over the Hypertransport connection. How much does this design boost performance? Quite a bit, according to our benchmarks. This relatively simple design decision offers a 6% performance increase.

The Iwill board is a tiny bit slower than the MSI board, and that might raise some eyebrows. However, Vtune tells us that the Xeon Nocona (1 MB L2) needs to access the RAM memory only 2% of the time. Assuming that the Opteron with its 1 MB cache needs about the same, it is clear that memory bandwidth is not going to determine the results by much. Slightly more aggressive timings (and thus lower latency) or clock speeds might give MSI the edge. These tiny performance differences are not important, however.

Benchmarks MySQL: Hyperthreading?

What can hyperthreading do for MySQL performance?

Concurrency Dual Xeon (Irwindale)
3.6GHz with HT
Dual Xeon (Irwindale)
3.6GHz no HT
HT On vs HT off
1 286 287 0%
2 450 457 -2%
5 497 559 -11%
10 517 583 -11%
20 545 561 -3%
35 506 573 -12%
50 495 570 -13%
       
AVG 512 569 -10%
Max 545 583 -7%

Amazingly, Hyperthreading decreases performance by quite a bit. This leads to a rather weird conclusion. If you want maximum MySQL (Read) performance from your Xeon server, you have to disable Hyperthreading and run in 32 bit mode. The former is of course not dramatic. The latter might, in some cases, be a serious limitation.

Benchmarks MySQL InnoDB: Intel versus AMD

What if we change the MyISAM engine for the ACID compliant, row level locking InnoDB engine under the hood of MySQL? Surely that should make scaling better, as the MyISAM table locking mechanism is simple, but could be one of the reasons why it scales less in multi-CPU configurations. Let us take a look.

Concurrency Dual Xeon (Irwindale)
3.6GHz with HT
with InnoDB
Single Xeon (Irwindale)
3.6GHz with HT
with InnoDB
Dual Xeon (Irwindale)
3.6GHz without HT
with InnoDB
Dual Opteron 248
Dual Channel

With InnoDB
Single Opteron 248
Dual Channel

With InnoDB
1 207 191 210 216 192
2 283 201 303 312 223
5 324 219 334 396 259
10 319 204 360 397 242
20 301 199 330 357 236
35 281 193 308 353 221
50 274 181 298 333 209
           
AVG 300 199 326 366 233
MAX 324 219 360 397 259

The InnoDB engine is at about 60% of the speed of the MyISAM engine. Let us analyze these numbers in detail.

Concurrency Dual versus Single Xeon Dual versus Single Opteron Dual Opteron vs Dual Xeon HT on vs off
1 8% 13% 3% -2%
2 41% 40% 3% -6%
5 48% 53% 19% -3%
10 57% 64% 10% -11%
20 51% 51% 8% -9%
35 45% 60% 15% -9%
50 51% 59% 12% -8%
         
AVG 51% 57% 13% -8%

Yes, we only used the 2.2 GHz Opteron 248, due to time constraints. We tested with this CPU because we also tried to get some numbers on the Dual core Opteron 275 (also 2.2 GHz), but as you know, we could not get that CPU running at dual core in SUSE SLES 9 SP1. It is pretty clear that a 2.6 GHz Opteron 252 would bring in another 16% - 18%. So, even with a different engine, the Opteron keeps outperforming the Xeon with a significant margin. This margin can again be lowered by disabling Hyperthreading.

The Opteron scales a little better than the Xeon in this test. All in all, the InnoDB scales better than the MyISAM engine, but not spectacular: a second CPU offers a 50% - 57% boost instead of 40% - 41% one.

What happens if we use the Dual core Opteron 275? To make this work, we had to resort to the Gentoo distribution again, with the 2.6.12 kernel. All CPUs are running at 2.2 GHz.

Concurrency Dual Dual Core 875 Single Dual Core 875 Dual Opteron 248 Dual Dual core vs One Dual core Dual core vs Dual single
1 199 206 200 -3% 3%
2 308 305 293 1% 4%
5 397 368 338 8% 9%
10 401 379 345 6% 10%
20 400 359 308 11% 17%
35 388 342 305 14% 12%
50 361 322 290 12% 11%
AVG 389 354 317 10% 12%
MAX 401 379 345

InnoDB does not scale better with 4 cores than MyISAM. On the contrary, both Engines show very small performance benefits from more than 2 cores. Interestingly once again, the dual core CPU is quite a bit faster than our Dual CPU (single core) machine. A 10% bonus is nothing to sneeze at, especially when you consider that server boards with only one socket are quite a bit cheaper. It seems that one dual core Opteron is an ideal solution for a rather powerful MySQL database server.

Next, we test with an enterprise database solution: DB2 8.2.

Benchmarks Benchmarks (continued)
POST A COMMENT

45 Comments

View All Comments

  • linuxnizer - Tuesday, July 19, 2005 - link

    Late contribution...

    The article mentions that Linux didn't work well with AMD Dual Core. The reason could be this:

    http://www.iwill.net/inews.asp?n_id=35

    it says:

    NVIDIA CK804 does not support dual core under Linux yet, only under Microsoft Windows.
    Reply
  • Illissius - Saturday, June 18, 2005 - link

    Nice article. I'd also be interested in PostgreSQL, being the "other" major open source database... specifically, whether it's any better at scaling with multiple CPUs. (Not that I have any practical use for this information, I'm just curious.) Reply
  • Viditor - Saturday, June 18, 2005 - link

    Seriously, mickyb and elmo may be correct about the Intel compilers (I frankly don't have a clue what's used in most shops)...

    The real problem is that it's a virtual impossibilty to create a "level playing field", but I have to say to the critiques of the article that Johan has done a stellar job of coming as close as possible!
    Reply
  • Viditor - Saturday, June 18, 2005 - link

    "They aren't testing compilers"

    Oh sure...just throw REALITY into the mix why don't you...!
    ;-)
    Reply
  • Icehawk - Saturday, June 18, 2005 - link

    They aren't testing compilers. Reply
  • Viditor - Saturday, June 18, 2005 - link

    mickyb - Thanks for the input! Fair enough...maybe Johan could use both the Pathscale compiler (which is optimized highly for Opteron) and the Intel optimized compiler on his next series of tests? Reply
  • mickyb - Saturday, June 18, 2005 - link

    I dissagree with the comment that a large number of people don't use the Intel compiler. I (other developers and IT shops) only use Intel compiler's for Linux. It is the fastest one out there for x86 and Itanium.

    If you are running a large database that requires a large server (compared with a desktop loaded with RAM to run a personal blog site) like this article is testing, you will be setting up the environment with a trained IT professional that will use the compiler that is fast and stable.

    When we build our product for all the UNIX platforms, we always use the vendor compiler instead of gnu. gnu works great and is free, but it is not optimized nearly as much.

    This is like saying the same audience won't recompile Linux on the platform they are going to install it on. This is the first thing you should do....and with an Intel compiler. There should be no real reason why one vendor Linux is faster than the others except for compile options and loaded modules. You cannot run Linux out of the box, it doesn't come in a box where I get it. :)
    Reply
  • DonPMitchell - Saturday, June 18, 2005 - link

    We need to see TPC-C benchmark results for MySQL and other new database systems. Why won't they step up and allow themselves to be compared to the major commercial systems?
    Reply
  • Viditor - Saturday, June 18, 2005 - link

    ElMoIsEviL - "as much as I am un-biased"

    C'mon mate...anybody who has read your posts knows you're heavily biased towards Intel, just as people who have read mine know that I am biased towards AMD. The important thing is to try and set aside the bias to look at things from both sides...I do try, but admittedly don't ALWAYS succeed. :-)

    I imagine you probably posted before you read the explanation of what a query cache is...understandable.

    As to not using an Intel specific compiler, I suppose that if it HAD been used I would be complaining as well. We have to rely on Johan and Anand (who frankly know a Hell of a lot more about this than either of us) to choose based on what the market actually uses...if you can site impartial industry sources that show otherwise, I'm sure we would all (especially the AT staff) would love to see them.
    I do know that over the years, Johan and Anand have shown themselves to be quite unbiased in their articles (you should go read some of them on Aces as well)!

    There are certainly things that I could pick apart as well..e.g. when he states
    "In the second half of 2004, already one million EM64T Xeons were shipped"
    Yes they were shipped, but that doesn't mean they were sold. The majority of those shipments were probably to OEMs for inventory buildup. Remember that Intel had a huge inventory write-off at the same time, and this was most likely a shift in inventory.

    Regardless, none of this has to do with the validity of the article which is excellent and makes sense. If you think about it, it should have been expected...the only for AMD to have increased their marketshare in servers is by performance. They certainly don't have the budget or marketing clout that Intel has!
    Reply
  • S390 guy - Saturday, June 18, 2005 - link

    About ISAM and DB/2... ISAM (Indexed Sequential Access Method) is NOT a database! It has no referential integrity nor rollback/commit features (although those can be activated on mainframe). ISAM was popular on mainframe when there wasn't any database (or rather when database was a too massive application to run!) and even there they were superseeded by VSAM. They're not much different from DOS random access files (an index file pointing to the relative record number on the main file).
    And it's no suprise that DB/2 scales well: mainframes rarely feature a single CPU, at least as far as I know.... IBM have had some 20 years to practice on multi-cpu machines!
    Reply

Log in

Don't have an account? Sign up now