Sun's W2100z Dual Opteron Workstation
by Kristopher Kubicki on October 27, 2004 12:05 AM EST- Posted in
- Systems
The Test
Testing the w2100z is difficult, since there are very few dual Opteron 250 rigs on the market right now. We had the opportunity to set up a dual Opteron 250 workstation using store-bought components for our tests. This will likely change as more Tier 1 workstation companies start working closer with AMD. The majority of our tests are conducted in the same manner as our Linux CPU and GPU tests in the past. Since this is a workstation, we want to focus the majority of our benchmarks on applications that we would use in a workstation: compiling, rendering and encoding tests.We took various benchmarks that we are familiar with and ran them on various configurations of the w2001z. As we mentioned earlier, the w2100z does not come pre-installed with any operating system. As a result, we took several different operating systems (Solaris 10, JDS 2.0, RedHat 9, SuSE 9.1) with similar gcc/kernels and benchmarked our programs on them. For reference, we took our off-the-shelf Opteron 250 rig and ran the same benchmarks on it using SuSE 9.1 as well. Although we upgraded the GCC libraries and glibc for the JDS 2.0 configuration, we did not change the kernel - JDS was designed around the 32-bit SMP Linux 2.4 kernel, and upgrading/hacking it would put us well outside the scope of the analysis. As we will see in the next few pages, performance on the proven 2.4 kernel is actually surprisingly good!
Finding properly threaded applications on Linux that are capable of taking advantage of multiple CPUs is a difficult task. While lots of server applications are designed for multiple threads, as the need for multiple users exist, applications that fully utilize multiple processors in a single user environment are difficult to come by. This makes it very challenging for us to benchmark applications effectively using multiple processors. Almost all of our benchmarks only utilize a single CPU unless stated otherwise.
All of our tests take place in the 64-bit environment, with the exception of the JDS rig. Although there is nothing to prevent us from running on 32-bit operating systems, (most) of our applications are mature enough to take a performance hit in a 64-bit environment.
| Performance Workstation Configurations | ||
| Processor(s): | (2) AMD Opteron 250 (130nm, 1MB L2, Socket 94) | |
| RAM: | 4 x 1024MB Buffered ECC PC-3200 CL3 (400MHz) | |
| Motherboards: | Sun K85AE | Tyan K8W S2885ANRF |
| Hard Drives | SCSI u320 Seagate Cheetah 10,000RPM | SCSI u320 Quantum Atlas 10,000RPM |
| Memory Timings: | Default | |
| Video Card(s): | GeForce QuadroFX 3000 | |
| Operating System(s): | SuSE 9.1 Professional RedHat 9 JDS 2.0 Solaris 10 |
SuSE 9.1 Professional |
| Kernel: | Linux 2.6.8 Linux 2.4 (JDS 2.0) SunOS 5.10 s10_63 (Solaris 10) |
|
| Compiler: | linux:~ # gcc -v Reading specs from /usr/local/lib/gcc/i686-pc-linux-gnu/3.4.2/specs Configured with: ./configure Thread model: posix gcc version 3.4.2 |
|
| Drivers: | NVIDIA Linux 1.0-6111 | |
The majority of the differences between these two systems result from different motherboards and slightly different hard drives. Most of our benchmarks do not particularly stress hard drive IO, and focus more on CPU and memory IO. Generally, we average three passes of our results unless stated otherwise. Variance is noted, if above 2%.
Facebook
Twitter
RSS
47 Comments
View All Comments
t - Thursday, October 28, 2004 - link
oh... i better clarify before i get labelled as a 'zealot' or a 'mac hater' or a 'pc lover'by 'cache starved' i mean that the power4 is _very_ dependent upon its cache architecture, take some of that away and u of course impact performance... a power4 and a G5 at the same clockspeed, the power4 wins. The G5 is still an impressive chip.
t.
t - Thursday, October 28, 2004 - link
heh...this thread is hilarious...can u ppl like talk past each other some more?please :)
G5 = cut down, somewhat cache starved power4
Blue Gene/L = power4+
they are a fair bit different: l3, l2, altivec, dual v. single core...just for a couple.
gromm - Wednesday, October 27, 2004 - link
"and I'm sure that there's nothing special about the way the G5's were linked... "Actually, they have a communication network based on InfiniBand, which isn't something that you'd buy for home (especially considering how much it costs). The cards themselves are $200+ each (in quantities of 10,000 for the only price I could find for their HCA cards) and I can't even find how much the switches cost (I'm sure several thousand dollars each).
Reflex - Wednesday, October 27, 2004 - link
Thanks for the correction, its been a while since I read up on that stuff. However you still illustrated my point that this is pretty much an irrelevant benchmark for general purpose computing. People do not simply use thier PC's for floating point performance...slashbinslashbash - Wednesday, October 27, 2004 - link
Reflex: "First off, once again, you are misunderstanding what you are looking at. Total number of CPU's is only part of the equation. There are *many* factors that go into the 'most powerful supercomputer' equation. How much memory and what type/speed? How are they linking the individual nodes? What kind of software optimizations have been done, and what software is being used to benchmark it?"Wrong. The top supercomputers are rated solely by FLOPS (Floating Point Operations Per Second, as I'm sure you know) as measured by the Linpack benchmark. See www.top500.org. I've never heard of memory having an impact on FLOPS; I guess it *could* if you absolutely starved the CPUs of work, but presumably all of these computers are balanced enough that the RAM can keep up with the CPUs. The nodes can be linked in any way; presumably they're linked optimally for price/performance, and I'm sure that there's nothing special about the way the G5's were linked... you don't build a supercomputing cluster and let the linking drag the performance down. As for the optimization question, I'm not sure but I'll bet Linpack is optimized for every platform/architecture.
Reflex - Wednesday, October 27, 2004 - link
I have not been arguing about superiority for the PC platform. My point is that they are not directly comparable as related to this particuliar review. The product being reviewed does not have an equivilent on the Mac side of things, so going on about how this article proves that the 'price' arguement is wrong is rediculous.My original question has not been answered, and that is that I am wondering who is building these since I have an identical workstation here on my lab bench but with an IBM label on it.
karlreading - Wednesday, October 27, 2004 - link
enuff of the mac vs. pc B$ dudes!!!this is a comments section about a opteron workstation, not about how a g5 spanks / get spanked by opterons ass.
that said there is one part of this that gets me excited. Whilst coming across mac / pc arguments on forums, i have noticed one trend. AMD is now always the PC's defender. i never hear anybody citing the p4 / xeon as a mac comparison. its always opteron / a64 vs. g5. this is excellent news from AMD's standpoint, as it cements the trend that AMD is a respectable company, and also is impressive to see AMD as the lead PC saviour in the ongoing annoying "my pc's better than your mac " debate. Intel should be worried, very very worried. never thaught id see the day when my beloved AMD were championing the pc / x86 cause :)
gromm - Wednesday, October 27, 2004 - link
As far as cost, I'd like to see how much Apple (and other sponsors) subsidized X. The networking infrastructure it has alone would normally be massively expensive and I can't see how it fits into the $10M pricetag quoted by them along with all compute hardware.From benchmarks I have run, the G5 and the Athlon64 are neck-to-neck in performance (actually the difference is small enough to be noise) in 'normal' codes (mostly FPU) and the Athlons are a little faster in integer performance. I haven't seen what Altivec/SSE2 optimizations would do for either.
If you want some rough estimates, go to Ars Technica and look in the Battlefront forum under the Cinebench thread. There are lots of scores under there to compare for this benchmark (includes a raytracer and some other stuff).
michael2k - Wednesday, October 27, 2004 - link
I'd like to see Anand run SpecViewPerf on his Dual G5, now :)Reflex - Wednesday, October 27, 2004 - link
I am aware of Apple's 'server' aspirations. That does not change anything at all. They do not have the kind of corporate support Sun or other large venders provide, and as a result the Xserve is not a large player in the market. Furthermore, its only proof that your comparing the wrong product to the Sun product this review was about. Xserve was designed to compete with workstations like this, not the PowerMac which is a desktop system.My comment about Sun relates to the fact that for a long time they were a detriment to the industry at a whole, pushing concepts like Java PC's with no local storage, trying to keep prices very high, and generally siding with 'Big Iron' in the market rather than embracing the future. In the past year, as Microsoft/Intel/AMD have made the Sparc obsolete they have had to get with the program, choosing AMD as their partner made perfect sense as they had no motivation to improve Intel's position considering they are still competing with them in some markets.
In summation, Sun is finally seeing the light, however their past is one of high prices, legal shennanigans(especially in Europe where price fixing has been a common charge against them), and a strategy of defining themselves as Microsoft/Intel's opposition rather than charging their own course. The future will tell where they go, and I'll cross my fingers and hope that x86 Solaris and Opteron workstations are a sign that they are finally producing products their customers demand, rather than locking them into a model and then telling them what they need...