I’m in the process of attempting to benchmark a multithreaded application on my new HP Proliant Server which has an 2x Opteron 6272 and 64gb of ram.

When I run the application on a desktop machine (a range of i7s and a Xeon X5675 processors) the application will cause all cores to hit a near 100% utilization.

When I run the application on my server, no matter how many threads I run, the total cpu utilization of the application hovers around 20-25%. That is if I’m running with 32 threads, all 32 cores will hang at around 20%, if I run 16 threads they’ll hang around 40%, and so on.

1. At first I suspected this had to do with the operating system, so I
   installed Windows 7 on the server so that the desktops and the
   server had the same OS.
2. Then I suspected it was the hardware, I changed the power management in the bios to High Performance. Even though this did increase the benchmark time, the same 20% utilization problem persists.
3. I can get all 32 cores at 100% using the y-cruncher benchmark. My custom benchmark is written in .NET, could this possibly have anything to do with it?

I’m perplexed by this problem. Anyone have an idea of what could cause this?

If your app is processing a large amount of data, try following the data’s path – if the input data is fed from the network, check for possible latencies, bandwidth limitations or transmission errors. You already checked disk I/O which otherwise would be a likely candidate for a bottleneck.

Last but not least, since it is a highly multithreaded .NET application, you should make sure that server garbage collection is used, otherwise you might see weird load characteristics as described in this post from stack overflow.

Check more discussion of this question.

I currently rent a server:

Intel® Xeon® X3440 Quadcore 4x 2.1 GHz, 16 GB DDR3 ECC

Now I have a possibility to get the following server for a cheaper price:

AMD Opteron™ 3280 Octo-Core 8x 2.4 GHz, 24 GB DIMM (DDR3) ECC

Do you reccomend me to move to a new server? I don’t know a lot about differences beween Xeon and Opetron. Thanks.

Firstly do you NEED extra performance, as moving from one to the other will be time consuming, risky to some degree and incur an outage.

That Xeon has hyperthreading, so you can kind of think of it as being able to do eight things at once, same as the Opteron. Normally I’d always take an Intel over an AMD as on anything but price Intel almost always rules the roost but there’s at least one generational leap between the 2009 Xeon and the much modern Opteron – and that does cancel out the ‘Intel lead’ I’d normally go with.

One thing’s for sure, unless you’re memory-starved or use code that hates hyperthreading then there’s not that much in it between them, though if this were a new system I’d probably go with the Opteron (or nag the Xeon supplier for a more modern chip).

But it really does all come down to whether the ~0%-20% overall likely performance difference is worth the move.

Check more discussion of this question.

howto configure memory around the Xeon 56xx CPUs?
2P configuration with 3 channels per CPU

How to fill the DIMM slots for optimum memory performance?
what is the minimum number of slots to be filled ?

dual channel, dual rank etc?

For the 5600, the max performance comes from populating one DIMM per channel and distributing DIMMs evenly across CPUs. So, 3 DIMMs per CPU in your case. That would also mean one DIMM per channel (e.g. slots 1, 4 and 7, starting from the slots nearest the CPU)

Check more discussion of this question.

I tried looking at the specs but they seem very similar, and I’m not sure which is the more powerful or feature rich chip?

Scalability is the main difference. The 270 can scale to 2 way while the 1212 can go to 1. the leading digit determines the scalability. That seems to me to be the main difference from what I can see. This is given if you were looking at the two models with similar clocks and cache.
Made a little comparison chart for you click here if you want it in its own window

Check more discussion of this question.

I am building a server which will host various websites on IIS, and also have multiple MSSQL databases. I’m currently trying to decide between 2 x AMD Opteron 6128′s (8 cores each) and 2 x Xeon E5620′s (4 cores each + Hyperthreading).

What do you think would be the better choice, and why? I can’t find too many benchmarks of those two in a web application scenario like mine. Thanks for your help!

What do you think would be the better
choice, and why?

The one with more discs. Seriously.

Put in enough memory and your websites will have to be very special do even start using the processor. I remember serving 400.000 visitors per hour on a dating site I made for a customer during a peak (tv appearance) with a dual pentium.

BUT: multiple sql server databases will use a lot of IO. This is a LOT more important than focusing on CPU because it is a lot more likely bottleneck. So, take the one where you can put in 10 or 12 or more drives and that is going to give you a lot better SQL performance. With fast discs.

Check more discussion of this question.

I’m currently working with an AMD Opteron-based NUMA system. For the needs of my current project, I’d like to make Linux and all of the system processes to utilize only the CPU0 (and preferably, only one of its cores), leaving all other cores for my own egoistic needs. I know that it will really impact system performance, but it is OK now.

So, I’d like to know if it is possible to achieve at all without messing with the kernel source code. Any links to the up-to-date articles about NUMA and SMP implementation details in Linux will be appreciated, too.

leaving all other cores for…

That implies that you want to actually use the other cores.

Before you start using using the other cores, use taskset to apply the affinity for all running user processes (including init). e.g.

taskset 0x00000001 1

Then set the affinity mask to everything else for the process which will launch your “egotistical needs”, e.g.

taskset 0xFFFFFFFE $$

You can’t force the kernel to run on only one CPU (and it would be stupid anyway) unless you set the boot options which will only allow the system to access a single CPU.

Check more discussion of this question.

i am trying to build a server for one of my customer.

the hardware is a HP ProLiant DL165, with an 8 core AMD Opteron 6128 and 4GB RAM. the server MUST run under Debian 5 (Lenny) due to a software dependency.

the standard Debian kernel does not support the 8 cores of the opteron and i am forced to use the nosmp boot option, which renders the processor quite useless by using only 1 core…

i read somewhere that i need a bigsmp kernel in order for my distro to support the 8 cores. is it true ? where can i find a bigsmp kernel for Debian 5 ? if i can’t find a ready-made one, how do i build a bigsmp kernel ? which config options are relevant ?

A 64-bit kernel shouldn’t have this problem. It will support all 8 cores. If you’re stuck with 32-bit for some reason, the BigSMP kernel is what you’ll need to make it work.

Check more discussion of this question.

I’m doing some bioinformatics work where the limitation is amount of memory (i.e., the process is not parallelizable). I see servers like the Dell R715 that have two sockets for Opteron 6100 CPUs and 128GB of RAM possible with 16 x 8GB DIMMs.

Each CPU has a bank of DIMMs associated with it.

My question is, if I put Linux on a machine like that, is only half the RAM addressable by each CPU? In other words, if I run a single-threaded program under Linux will that process have access to 128GB of RAM or only 64GB of RAM?

Strictly speaking, that single-threaded process will have access to all memory available to the server. You don’t have to worry about that.

The process will have access to all memory in RAM, but access to half of it will be (in theory) slower than access to the other half of it. I know VMWare-ESX is aware of memory locality, as it attempts to keep all the memory for a specific VM on the same memory node as the process execution in order to keep performance good.

The technology is called Non-Uniform Memory Access (NUMA). Chances are your Linux machine already has the bits for it (numastat should return data). The numactl program assigns a memory policy to your processes, which can be very useful if you’re writing your own code and want to optimize for memory speed. CPU-local memory will fetch faster than memory in another CPU’s memory node, which may be useful for you depending on what you’re doing with it.

Check more discussion of this question.