Sign in with
Sign up | Sign in
Your question

Powerful Server needed

Last response: in Components
Share
November 11, 2010 10:31:01 AM

Hi,

I'm looking for a system with following specs:
* 128 GB RAM or above, 8 socket Multi-Core (basically max cores)

Should I go for
1. A single 8-Socket machine OR
2. 8-Single socket machines in distributed mode OR
3. A mainframe

The applications run on it are distributed computing enabled. Although availability is paramount, I would rather demote it in this decision. I'm looking for other architectural bottlenecks, performance as key decision-drivers.

Please suggest

Regards,
CK

More about : powerful server needed

November 13, 2010 11:27:37 PM

ckinmum said:
Hi,

I'm looking for a system with following specs:
* 128 GB RAM or above, 8 socket Multi-Core (basically max cores)

Should I go for
1. A single 8-Socket machine OR
2. 8-Single socket machines in distributed mode OR
3. A mainframe

The applications run on it are distributed computing enabled. Although availability is paramount, I would rather demote it in this decision. I'm looking for other architectural bottlenecks, performance as key decision-drivers.

Please suggest

Regards,
CK


You're looking at some serious hardware costing a lot of money here. I would seriously consider asking somebody who is truly familiar with this line of work before dropping the tens or hundreds of thousands of dollars that some of the options on your list cost. I'll give you a little bit of insight as to what you may want to look for, though.

A single multi-socket machine's big advantages are that all data is accessible locally in a single image and also that on-motherboard interconnects are much higher-bandwidth and lower-latency than off-motherboard connections such as Ethernet, Infiniband, or other interconnects. Its big disadvantages are that it is extremely expensive for the throughput the CPUs are capable of and it is also a single point of failure. You're going to pay at least a few thousand dollars per socket just for the processors in an 8-socket system, for example. The only machines I know of that can be run with 8 or more processors are Intel's Xeon 7500 series, Intel Itaniums, IBM POWERs, and Fujitsu's SPARC units. All are extremely expensive compared to eight single-socket machines with CPUs of comparable performance. Do note that the price of servers goes up precipitously once you go beyond 4 socket as 8-socket and above machines are pretty rare and the vendors can thus charge out the nose for them.

Your application is distributed computing enabled, so you do not need a single memory image- otherwise the work would not be able to be distributed to remote machines. Thus the only reason you would really want to have one great big machine instead of a bunch of smaller ones is if the application creates an ungodly amount of inter-thread traffic that would swamp an off-motherboard communication mechanism. If you have a manageable amount of traffic, then get yourself a bunch of inexpensive two-socket server boards and link them together through Ethernet or Infiniband or whatever you would like to use. I'd recommend using two-socket server boards since almost all consumer motherboards do not support ECC RAM or headless operation, but server boards do. A dual-socket board doesn't cost very much more than a single-socket server board and there isn't a price premium for dual-socket-capable server CPUs compared to single-socket-capable ones (at least from AMD, who offers CPUs with much higher core counts and lower prices than Intel). You'll end up with a lower cost overall since you're buying half the number of cases and power supplies and such as well. If you are after a lot of cores, I'd look at the Opteron 6100 series since they come in 8 and 12-core variants and some can be had quite inexpensively. Those particular Opterons can be run in up to four-socket machines as well, and four-socket G34 motherboards run about twice what dual-socket ones do. However, going to a quad-socket machine can get pretty expensive since the boards require some unusually large and expensive cases that can be difficult to find.

So if I were you, I'd try to get together a couple of standard desktop machines and try to run the application in distributed mode and see how well it scales to see if you need to go the one huge machine route or the several smaller machines route.
m
0
l
November 16, 2010 4:17:41 AM

Thanks MU_Engineer.... In fact, I'm a software guy with a bit of knowledge on hardware - so this was enlightening !!

For a new application being developed for my employer, I was asked to draw up hardware requirements. It's just a simulation engine - which means, more than floating point operations, I'm looking at large data manipulation.

Strangely, today I saw that we could do data partitioning in such a way that each machine can perform independent processing. So I guess, multiple dual socket / single socket machines fits the bill. Although off-MoBo communication is expected to be slightly heavy, I don't expect it to swamp the network.

Due to employer's inclination towards Intel products, I don't have too many options unless I build a strong case otherwise.

As you said, a low cost dual socket multi-core Xeons (4 to 5 machines) running Linux would actually be better.
m
0
l
Related resources
November 16, 2010 4:09:37 PM

ckinmum said:
Thanks MU_Engineer.... In fact, I'm a software guy with a bit of knowledge on hardware - so this was enlightening !!

For a new application being developed for my employer, I was asked to draw up hardware requirements. It's just a simulation engine - which means, more than floating point operations, I'm looking at large data manipulation.

Strangely, today I saw that we could do data partitioning in such a way that each machine can perform independent processing. So I guess, multiple dual socket / single socket machines fits the bill. Although off-MoBo communication is expected to be slightly heavy, I don't expect it to swamp the network.


Glad I was able to help.

Quote:
Due to employer's inclination towards Intel products, I don't have too many options unless I build a strong case otherwise.

As you said, a low cost dual socket multi-core Xeons (4 to 5 machines) running Linux would actually be better.


Both Intel and AMD server CPUs are solid and each have their advantages and disadvantages based on specific usage scenarios. I'm more of a hardware guy, so I'll give you some information you and employer can look at to determine which ones may be better for your usage:

Advantages of current AMD Opterons over Intel Xeons
1. Considerably higher core counts for the same amount of money. For example, $800 will buy you either a quad-core Xeon E5640 or it will buy you a 12-core Opteron 6168.
2. Better overall performance in multithreaded tasks when comparing similarly-priced CPUs, sometimes considerably so.
3. Significantly lower prices for entry-level systems. A pair of the least-expensive Opteron 4100 quad-core CPUs cost $200, which is what Intel charges for one of its lowest-priced Xeon 5500-series CPU, and that Xeon is only a dual-core unit.
4. The lower-priced units have all of the same capabilities as the higher-priced units in the same product line. The only thing that changes is the amount of cores and the clock speeds. Memory speed, bus speed, and cache sizes, and features are all the same in a particular line of Opteron CPUs. The lower-priced Xeons have less cache, slower memory speeds, slower bus speeds, and some have hardware features like HyperThreading and Turbo Boost turned off as well.
5. Current Socket C32 Opteron 4100 and Socket G34 Opteron 6100 platforms can be upgraded to take the next generation of Opteron CPUs that will come out next year. The current LGA1156 Xeon 3400 and LGA1366 Xeon 3500/3600 and 5500/5600 platforms cannot be upgraded to the next generation of Xeons since they will use different sockets and platforms.

Advantages of current Intel Xeons over AMD Opterons
1. Somewhat better per-clock and per-core performance, so a Xeon with the same number of cores and same clock speed as an Opteron will be faster.
2. The high-end Xeons have much higher clock speeds than the Opterons, so programs that are not all that well multithreaded will run considerably better on the high-end Xeons versus high-end Opterons.
3. Intel's ICC compiler can handicap the performance of AMD processors. If you are running a proprietary program that is available only in binary form and that binary has been compiled with ICC, AMD processors will perform worse than you think they would given their relative performance in other applications. I don't know if this will be an issue for you, but keep it in mind.
4. The absolute highest-end Xeons can sometimes be a little faster in overall performance than the absolute highest-end Opterons.
5. "Nobody ever got fired for buying Intel." If they like Intel and you lobby for getting AMD equipment and then they do so because of your lobbying, ANY problems with the hardware will be YOUR responsibility. That's not to say don't get AMD hardware, but let them make the final decision based on the costs and benefits of the setups. Then they own the decision and any possible problems, not you.
m
0
l
November 30, 2010 9:20:15 AM

ckinmum,

If you are still watching, I would note to you that GPU's have programmable cores that would likely be perfect for your application. Look up OpenCL for more information. A single, or dual, processor machine might be all the power you need when you combine it with several high end graphics cards, and enable their use through programming.

Kremnari
m
0
l
!