Single Dual Processor System vs Seperate Systems

[rdsn]

While discussing possible server configurations for heavily threaded intensive applications, I started wondering what would be more cost efficient and have better performance per dollar, two separate systems running i7 3770 cpus or one dual processor system running Intel Xeon E5 2620s. Assume the two i7 machines cost $2200 and the dual processor cost about the same. What do you guys think

 

[cl-scott]

While I have nothing at all to back it up with, I would think that the single system would be cheaper to operate long term.

I justify that by saying that you would have less overall overhead in a single system. Even if you pair down these systems to just the essentials, you're not duplicating the electronics. I would not expect that the differences would be very large, but long term even a small savings will add up over time.

 

[Ken g6]

You're kind of comparing apples and oranges here. A more equal comparison would be two LGA2011 systems. As it is, I don't see why two presumably non-overclocked i7 machines should come close to $1100 each, unless you have large graphics cards in each one.

I suggest you spell out the requirements of your machine(s) better. [thread=80121]Using this template[/thread] would be helpful.

 

[rdsn]

You are correct, last I figured i7 3770 machines should only cost $950 each and that is with overbuilt components for reliability. However, the choice is between a dual cpu machine or prebuilt i7 3770 machines. Also, last time I checked i7 3770s were comparable to the xeons in terms of performance. These machines would be running heavy cpu intensive multithreaded calculations in a server environment, closest example I can give is video encoding 24/7.

 

[Cerb]

For a server, your only choice is a Xeon, whether the E5 2620 or some other. If AMD were more competitive, you'd have other reasonable options, but no ECC in servers that handle anything more than ephemeral network duties is just not responsible.

For dev/test boxes, the i7 has a light edge, on the face of it, but only about 15% with perfect scaling (unlikely). Also, the extra cache on the Xeon, and being able to store everything in a single computer's RAM and drive's, might help it out, assuming you aren't dealing with an easy ALU-bound problem (like scientific computations).

The dual-Xeon system would have nearly double the RAM bandwidth (8xDDR3-1333 v. 4xDDR3-1600), should that be a limiting factor.

Like has been said, though what you're doing with it matters, and it also seems like the i7 systems would be cheaper, as would 2 single Xeon servers.

 

[Kippa]

Out of interest if you had a could you use a dual xeon as a gaming rig as well as using it for pro applications. The only reason why I ask this is because I remember years ago there was a quote saying that the game supreme command had problems running on xeons, although I am not sure if that is true.

 

[rdsn]

For mission critical stuff sure, but the reason why i7s are being considered in the first place is because the servers would only be doing less important tasks that are relatively light. It won't be nice if the thing breaks down all the time but ECC is more of a benefit than a requirement, as is other server features, such as 64gb of RAM. The only thing that really matters is performance per cost.

@Kippa
Dual processors just show up as processors with enormous amounts of threads but there is no reason to game on dual processors because their single threaded performance is pretty horrible for the amount you pay

 

[Ken g6]

It really would have helped if you filled out at least some of the sticky thread I linked to. I get the general idea, but there are still several factors to consider.

First, do you want to overclock? It looks like perfect calculation every time isn't critical, so I'd consider it.

Have you measured how much hyperthreading benefits this application? If it's little to no benefit, dropping to 3570k's could save a couple hundred.

Is this application floating-point intensive? If it uses little to no FP, Bulldozer might actually be an option.

Do you happen to be using discrete video card(s) with these systems? Or do you have some old video cards lying around that you could use?

How much RAM, SSD space, and HDD space do these systems need? Any other add-on cards or unusual interfaces (e-sata, ieee1394, etc.) required? Non-processor components could greatly affect prices here.

At this point I see a handful of processor options:

• If overclocking:
  • 3770k
  • 3570k if hyper-threading isn't beneficial.
  • 3930k x2 on this board
• If not overclocking, a B75 board with one of:
  • 3570 if hyper-threading isn't beneficial.
  • Xeon E3-1230 V2 if you're using discrete video.
  • Xeon E3-1245 V2 otherwise for single-processor systems.
  • Or your Xeon E5 dual-processor system.
• Or a single or dual-processor Bulldozer system

 

[mfenn]

While discussing possible server configurations for heavily threaded intensive applications, I started wondering what would be more cost efficient and have better performance per dollar, two separate systems running i7 3770 cpus or one dual processor system running Intel Xeon E5 2620s. Assume the two i7 machines cost $2200 and the dual processor cost about the same. What do you guys think

We really need more information about what these machines will actually be doing to make a good determination.

In general though, the more tightly-coupled the computation is, the more penalty you pay by going outside of the box. You can mitigate the network penalty by getting a really fast, low-latency network like InfiniBand, but that kind of defeats the purpose of getting cheap systems in the first place.

Another thing to consider is software licensing costs. If you use any commercial software for this application, licensing it for two machines will probably instantly destroy any cost savings.

 

[rdsn]

We really need more information about what these machines will actually be doing to make a good determination.

In general though, the more tightly-coupled the computation is, the more penalty you pay by going outside of the box. You can mitigate the network penalty by getting a really fast, low-latency network like InfiniBand, but that kind of defeats the purpose of getting cheap systems in the first place.

Another thing to consider is software licensing costs. If you use any commercial software for this application, licensing it for two machines will probably instantly destroy any cost savings.

They will be running custom programs that perform load tests on systems then store and analyze the results.

 

[Ken g6]

They will be running custom programs that perform load tests on systems then store and analyze the results.

You keep teasing with little bits of information, but not answering our questions. I had several questions you didn't answer:

1. First, do you want to overclock?
2. Have you measured how much hyperthreading benefits this application?
3. Is this application floating-point intensive?
4. Do you happen to be using discrete video card(s) with these systems? Or do you have some old video cards lying around that you could use? [note: most server systems, including all the dual-processor systems, I think, do require a discrete video card.]
5. How much RAM, SSD space, and HDD space do these systems need?
6. Any other add-on cards or unusual interfaces (e-sata, ieee1394, etc.) required?

Mfenn also asked about software licensing costs.

Now I'm intrigued, though. Are you making virtual user generators for LoadRunner, LoadUI, or some similar system? That would seem to be limited by Ethernet bandwidth, although I could be wrong.

 

[mfenn]

You keep teasing with little bits of information, but not answering our questions. I had several questions you didn't answer:

:thumbsup: Baby steps in the right direction here, but we still aren't getting much meat. I tend to agree with you though, if it's the load testing client, the machine(s) probably don't need much CPU horsepower at all for the actual load test.

Whether or not they need a lot of power for analysis really depends on what sort of analysis is going to be happening. We also don't even know if this analysis is something that can even be split across two machines in the first place.