Eight Core Servers: Opteron 880 (Egypt) vs. Xeon MP 3.0 GHz (Paxville)
by Jason Clark on April 24, 2006 2:00 AM EST- Posted in
- IT Computing
Software Configuration
Hardware Configuration
Benchmark Configuration
Many of you will notice that we've finally transitioned our benchmark platform to 64-bit. It's been a long road getting there; we were constantly trying out builds of SQL 2005 64-bit as it was in development, and had mixed results. The final build produced viable numbers, and we've fully transitioned to SQL 2005 for all future benchmarks.
Our goal throughout the tests was a constant CPU usage of 80-90%; we didn't want to reach the point of system saturation, nor did we want the systems to have too many idle cycles. Each test was first verified and then run 3 times. The standard deviation for the tests is 3-4%. We then averaged the set of three runs to produce the final result.
Linux MySQL Tests
SysBench 0.4.3 was used for the MySQL tests. We ran 4 different tests for each platform. The first two tests consisted of accessing 1M rows in a read-write scenario, and 1M rows in a read-only scenario. The second tests were the same only with 10M rows.
MSSQL Forum Test
The forum benchmark is a standard mid-tier OLTP test, which was made by replicating live query data from our forum database during peak hours. We took those queries and then record them in an XML file, with random row ID generators to handle keeping the test as real-world as possible. We wrote a C# application which takes the test and plays it back using several threads to stress the database to a desired level.
Dell DVD Store
The DVD Store Version 2 (DS2) is a complete online e-commerce test application, with a backend database component, a web application layer, and driver programs. The goal in designing the database component as well as the mid-tier application was to utilize many advanced database features (transactions, stored procedures, triggers, referential integrity) while keeping the database easy to install and understand. The DS2 workload may be used to test databases or as a stress tool for any purpose. The test is maintained by Dave Jaffe and Todd Muirhead from Dell. We configured the test to hit the database directly, instead of using a web-tier; we're testing database performance, and that removes any unneeded complexity and possible discrepancies.
Windows Tests
Windows 2003 Enterprise 64-bit was used for the tests along with SQL Server 2005 64-bit.
Linux Tests
CentOS 4.2 x86_64 was used along with MySQL 5.0.18 x86_64.
Hardware Configuration
Intel Paxville System
Quad 3.0 GHz Paxville Dual Core Processors (667Mhz FSB / 2x2MB L2)
E8500 (Twin Castle 4S) Xeon MP Truland Platform
16GB DDR2
8 x 18GB 15,000RPM Ultra320 SCSI drives in RAID-0
LSI Logic 320-2 SCSI Raid Controller
HP DL585 Opteron System
HP Proliant DL585 System
Quad Opteron 880 Dual Core Processors
16GB DDR
8 x 18GB 15,000RPM Ultra320 SCSI drives in RAID-0
LSI Logic 320-2 SCSI Raid Controller
Benchmark Configuration
Many of you will notice that we've finally transitioned our benchmark platform to 64-bit. It's been a long road getting there; we were constantly trying out builds of SQL 2005 64-bit as it was in development, and had mixed results. The final build produced viable numbers, and we've fully transitioned to SQL 2005 for all future benchmarks.
Our goal throughout the tests was a constant CPU usage of 80-90%; we didn't want to reach the point of system saturation, nor did we want the systems to have too many idle cycles. Each test was first verified and then run 3 times. The standard deviation for the tests is 3-4%. We then averaged the set of three runs to produce the final result.
Linux MySQL Tests
SysBench 0.4.3 was used for the MySQL tests. We ran 4 different tests for each platform. The first two tests consisted of accessing 1M rows in a read-write scenario, and 1M rows in a read-only scenario. The second tests were the same only with 10M rows.
MSSQL Forum Test
The forum benchmark is a standard mid-tier OLTP test, which was made by replicating live query data from our forum database during peak hours. We took those queries and then record them in an XML file, with random row ID generators to handle keeping the test as real-world as possible. We wrote a C# application which takes the test and plays it back using several threads to stress the database to a desired level.
Dell DVD Store
The DVD Store Version 2 (DS2) is a complete online e-commerce test application, with a backend database component, a web application layer, and driver programs. The goal in designing the database component as well as the mid-tier application was to utilize many advanced database features (transactions, stored procedures, triggers, referential integrity) while keeping the database easy to install and understand. The DS2 workload may be used to test databases or as a stress tool for any purpose. The test is maintained by Dave Jaffe and Todd Muirhead from Dell. We configured the test to hit the database directly, instead of using a web-tier; we're testing database performance, and that removes any unneeded complexity and possible discrepancies.
Facebook
Twitter
RSS
35 Comments
View All Comments
coldpower27 - Monday, April 24, 2006 - link
At least then by the time you do replace your machines hopefully Intel will be more competitive in this spaceAnemone - Monday, April 24, 2006 - link
Honestly it doesn't hurt to bring the point home again and again about Opteron. Those making buying decisions who are still stuck on Intel only solutions need the truth laid out for them regularly and repeatedly, imo. Thank you for doing that, because the current Intel solutions are a waste of money, and I hate that the most of all.:)
johnzo - Monday, April 24, 2006 - link
Are your test result comments correct? ie DVD Store Test, should the AMD lead be 40% not 29% ?AMD has 8853 more orders, which is 40% more than 21782.The same error (?) goes for the other percentages mentioned.
If I am wrong please forgive somebody with an old brain !
Jason Clark - Monday, April 24, 2006 - link
You are correct, my mistake. This is fixed.psychobriggsy - Monday, April 24, 2006 - link
Right, using funky-maths, if a 2.6GHz Conroe is 20% faster than a 2.8GHz X2.And this 8-core 2.6GHz (?) Opteron system is 40% faster than a 3GHz 8-core Xeon.
And the issue is scalability, i.e., the same problem will exist for Woodcrest.
However Woodcrest will be 3GHz, but this will only be useful for cache-bound tasks because of the scalability problems (expect all the Intel biased websites to only benchmark Woodcrest with this type of task). Maybe Opteron will be dual-core 3GHz by then...
Anyway I can't conclude anything without making huge vast leaps of guesswork, but yes, it will be very close later this year, but if Woodcrest can't beat Opteron because of the platform limitations then there will be little to no incentive to switch back from buying Opteron based servers to Intel based servers (for the companies that have switched to AMD), apart from power consumption.
However it means Intel will stop losing customers to AMD in the server space, if the performance and power consumption issues are addressed. By no means do I see AMD losing much of their gained marketshare however (and by Q4 it will be higher than now I'm sure).
peternelson - Monday, April 24, 2006 - link
If power consumption is your concern, then AMD offer reduced power versions of Opteron chips (eg the HE models) at increased cost. When they can migrate to 65nm process, we should see even better power economy.
Furen - Monday, April 24, 2006 - link
Here's what I think will happen. AMD will get a run for its money on the single-socket and dual-socket arenas (since dual-FSB pretty much keeps the FSB from being too limiting) but once you scale higher AMD will once again rule the roost. I also think that quad-core will be better on AMD's architecture, too, since throwing 4 Conroe/Woodcrest cores on a single FSB (266MHz I'd expect, though I suppose it could be 333) will give these cores a huge performance hit, more so in MP configurations (just thinking about the cache-coherency traffic for 8 cores scares me)coldpower27 - Monday, April 24, 2006 - link
Maybe, but it's hard to say at this point, how low FSB speed will affect Core Architecture, plus this is the Xeon MP space, where it won't be transistioning to Core Architecture for sometime until Q1 2007 earliest from what I believe with the Clovertown MP/Tigerton Core, I don't think there is a Woodcrest equivalent for this space.I expect AMD to have the "FSB" advantage in 4-8 way situations yes, however this may or may not translate to a performance crown for AMD.
Thoug I think it's pretty safe for Intel that they can take back the Xeon DP space with Woodcrest, and DIB.
JarredWalton - Monday, April 24, 2006 - link
I think you need to remember a couple of things. First, Woodcrest is based off of the next-generation Core architecture. Power and performance characteristics will be completely different from the current NetBurst chips. Second, Woodcrest will also have a 1333 MHz front side bus -- twice as fast as the server benchmarked here. Finally, and I could be wrong on this, but I think Intel is looking at one shared front side bus for every two sockets, and they could potentially move to one front side bus for each socket.That last idea would make sense for quad-core. And speaking of quad core, that's such a significant change that again it's almost impossible to predict at this stage. I mean, we don't even really know how AMD or Intel are going to build their quad core packages. Are we talking a single die, or will they have multiple dice on a package? Will Intel change the way the dice communicate with each other?
Of course, none of this things means that Intel will come out on top, but there are enough significant changes that we can't declare a victor at this point. Also, large L3 caches can indeed help server work. Otherwise, why would IBM even make a POWER5 chip with 128 MB of L3 cache? Adding tons of cache to a desktop system rarely helps, but enterprise servers are completely different beast.
At the very least, things should get interesting later this year. :-)
Furen - Monday, April 24, 2006 - link
You are indeed correct.In the DVD store test the AMD system is indeed 40% faster than the Intel system, but the Intel system is 30% slower than the AMD one. He seems to have chosen the AMD scores as the reference but incorrectly says that AMD is 30% faster instead of saying that Intel is 30% slower.