Intel 7500 Series "Nehalem-EX" Xeons
Posted on: 03/30/2010 10:57 PM

Usually when Intel launches a new platform I have a big box delivered to the house that contains a generic, white-box type "Intel Software Development Platform". Not that I'm complaining, but sometimes it's hard to get all fired-up over another plain, big, black tower sitting under my desk, regardless of what is inside. So imagine my surprise this time around when Intel told us that they had partnered with Dell for the Nehalem-EX test platforms!

We've all complained about "paper launches" in the past and having to wait months before the actual products were available (hence the generic test platforms) but that is not the case this time around. In fact, Dell had enough lead-time (i.e. had production silicon) to have final production machines ready to ship. I'm sure that Dell will have the new blades (M910) and 4U servers (R910) on their site this afternoon, but the most interesting one from my perspective is the 2U machine they sent me for evaluation.

Dell PowerEdge R810 Server

I won't lie, I am VERY familiar with Dell servers. I'm a small business consultant by day and over the years I have been responsible for the purchase and deployment of a LOT of servers for various clients. A good chunk of those servers have been Dell machines. My R810 came loaded with all the bells and whistles: PERC H700 RAID controller with 512mb cache, iDRAC6 Enterprise Embedded Managment Controller with 1gb SD memory, 1100w redundant PSUs, a pair of Xeon 7560 processors and 128gb of 1066mhz DDR3 (32x 4gb DIMMs). Needless to say, I was giddy like a school girl.

I really could write an entire article on the R810, but I'm going to try to hit all the highlights and still fit it on a single page.

Dell PowerEdge R810
Click image to view the gallery!

The PowerEdge R810 starts life as a 2U, 4-socket server with thirty-two... thirty-two... DIMM slots (2 banks of 4 DIMMs for each socket) supporting up to 512gb of total system memory! While on the subject of memory, let me mention the first little twist that Dell came up with for their Nehalem-EX platforms. FlexMem Bridge Technology.

FlexMem Bridge Technology was developed and owned exclusively by Dell. Basically, they came up with a QPI bridge that will allow their 4-socket Nehalem-EX machines to access all 32 DIMMs when running in a 2-socket configuration. So, you can buy the machine with only a pair of processors, but still load-up all 32 DIMM slots. That's a double bonus because it allows those on a tighter budget to save some money by buying lower density DIMMs while still getting the total memory capacity they need.

Apparently the engineers at Dell were thinking about people like my clients who are consolidating their small business networks to a virtualized environment. People like that aren't really worried about over-subscribing their CPUs, because a lot of their VMs use very little processing power. Where they are hitting the ceiling though, is allocating memory to the VMs. Now my clients can have their cake and eat it to!

Another small but ingenious thing that Dell did with the R810 to endear themselves to the virtualization market is to add mirrored SD memory slots in the server. So if you're booting your VMs from a SAN you can opt to get your R810 diskless and install the VMM of your choosing on a mirrored pair (think RAID1) of SD memory cards. They appear during setup as a normal drive which makes things REALLY easy.

Being a consultant, I've really come to appreciate things that make my life easier when it comes to managing my clients' machines. Embedded server managment controllers have become very important to me and the iDRAC6 does not disappoint. My server came with the upgraded "Enterprise" version of the controller which also adds the Vflash option. Vflash adds more memory, this time integrated on the iDRAC6 controller. This memory ships (but can also be user configured) with device drivers specific to your system so that you never have to hunt down the driver disk during provisioning (or re-provisioning) of the server again. You can litterally rack the server, pop in your OS install disk and then configure and deploy the machine without ever having to lay hands on it again.

I'm getting a little long winded here and it's time to move on but please feel free to browse through the R810 image gallery for an in-depth pictorial look at the server, as well as some screenshots of the remote managment interface and some other cool stuff.

The Comparative System?

When I was informed that I would be getting this machine to test I racked my brain trying to figure out what the heck I was going to compare it to. This is not generally the class of machine that I'm used to reviewing, so there is nothing kicking around the underground lair that would even come close to being similar. I've got 2-socket Nehalem and Westmere machines I could've compared to the Nehalem-EX platform, but the configurations aren't even close to similar, and the processors themselves are geared towards completely seperate markets.

I was all geared-up, though. All my workloads were in order, and I even had a brand new virtualization benchmark ready to rumble. All I needed was a similar machine to compare the Xeon 7500 platform to. To use a car analogy (for Sanjay and Greg, who apparently love car analogies), the Westmere-EP is really like a sportscar. If you're looking for the ultimate performance and you're willing to sacrifice a bit in other areas, by all means, go with the sportscar. But in the world of car analogies, Nehalem-EX is an SUV... and I needed another SUV to compare it to.

I was discussing my predicament with my buddy Paul Venezia over the weekend and he had a solution for me in the form of a quad Opteron box. While not identically configured to my R810, it was similar enough to make a valid comparison, so he went about installing Windows Server 2008 Enterprise Edition and setting-up remote access for me. I found my other SUV, but because of the time constraints and lack of physical access to the Opteron box, I had to make some adjustments to my usual workloads, and I had to cut out the power measurments and virtualization test completely. But, I think it was worth it, because the comparison is much more direct.

The Opteron box in question was equipped with four Instanbul based Opteron 8435 (6-cores, 9mb L3 cache and a 4.8GT/s HT link) processors, 48gb of DDR2 800 memory and a similar disk sub-system. Both test machines were loaded with Windows 2008 Server Enterprise Edition x64 with SP1 and all available Windows updates at the time of the testing.

Test Workloads

I used the following versions of our test applications:

- SiSoftware SANDRA 2010 SP1 Lite
- SunGard Adaptiv Credit Risk Analysis v4.0
- CINEBENCH R11.5 x64
- POV-Ray v3.7 beta 35a x64
- SPECjbb2005 v1.07
- Euler3d CFD Benchmark v2.2
- Matrix Multiply v3.5
- FlamMap (FSPRO)
- MyriMatch
- x264 Benchmark HD v3.0

Noticably absent from this review is an old-time favorite, 3ds Max. I did attempt to run our custom 3ds Max benchmark on both the 2009 and 2010 versions of the software, but the application would simply not load on any machine with more than 16 threads. Evidently Autodesk didn't plan far enough ahead to write their software for more than 16 threads. Once there is an update that addresses this issue, I will happily add 3ds Max back into the benchmarking mix.

Also missing from this round of tests is the Black & Scholes Kernel benchmark. It's been so long since I've tested an Opteron machine that my custom version compiled with the Microsoft compiler went missing. All versions supplied by Intel or compiled with their compiler failed to run on the Opteron machine, and due to time constraints I couldn't build a new version to use this time around.

All benchmark numbers presented were an average of three runs. Desktop resolution was set at 1920x1200, 32bit color with a 60Hz refresh rate. "Optimized Defaults" were used in the BIOS on all machines. 

Xeon processors were tested with SMT and Turbo set to ON for all tests. For the SPECjbb tests "hardware prefetch" and "adjacent cache line prefetch" options were disabled in the Xeon system's BIOS.

Printed from (,2.html)