2
I recently purchased a new work computer that I use mainly for scientific computation. Supposedly, it is a beast. Here are some relevant specs:
Processor 48x Intel(R) Xeon(R) CPU E5-2670 v3 @ 2.30GHz
Memory 131957MB (2047MB used)
Operating System Linux Mint 17.2 Rafaela
Kernel Linux 3.16.0-38-generic (x86_64)
It is replacing an old computer (bought in June 2011) with these specs:
Processor 8x Intel(R) Core(TM) i7-2600 CPU @ 3.40GHz
Memory 16390MB (7052MB used)
Operating System Linux Mint 13 Maya
Kernel Linux 3.2.0-23-generic (x86_64)
However, I noticed when I ran some simulations that I have been working on using the new computer that they are about 3x slower than they were in the old computer.
So I used hardinfo to benchmark. Here is what I got for the new system:
CPU Blowfish 0.507
CPU CryptoHash 1498.120
CPU Fibonacci 1.603
CPU N-Queens 0.688
FPU FFT 1.066
FPU Raytracing 6.421
And here are the results for the old system:
CPU Blowfish 2.070
CPU CryptoHash 582.869
CPU Fibonacci 1.471
CPU N-Queens 0.461
FPU FFT 0.860
FPU Raytracing 2.995
So the benchmarking also suggests that on certain dimensions the new machine is substantially slower than the old one. (I ran these tests a few times and the numbers are pretty stable.)
Any idea what could be causing this or how I can proceed with troubleshooting?
Both of these machines are Dells and I haven't changed the stock hardware or done any overclocking or anything like that. I also haven't messed around with BIOS settings on the new computer and I don't think I have on the old one.
And the simulations in question, do they deploy one processor core or several; is the software multi-threaded? – Run CMD – 2015-12-01T17:26:40.130
Note that if the benchmarker is single threaded, then the differance is likely the result of the frequency reduction (3.4GHz vs 2.3GHz or about a 30% drop). For instance the N-Queens problem is inheretnly single threaded because it is recursive and must compute the next state based on the previous. More CPUs/Cores will not speed these kinds of algorithms. but you can solve 48 n-queens problems at one time, whereas the previous system could only do 8. you want to up frequency to do one thing fast, and up cores/CPUs to do multiple things at once. – Frank Thomas – 2015-12-01T17:38:17.730
I understand your fundamental question I just don't understand the problem. Only a single benchmark is actual worse, unless I am reading the number wrong, and a smaller number is actually better. How do you figure the new system is 3x slower compared to the old system. I bet you anything
CPU Blowfish
is a single threaded. – Ramhound – 2015-12-01T17:45:33.990They involve solving a large number of relatively small linear programs using CPLEX interfaced with AMPL and Matlab. CPLEX is multi-threaded and is using all 32 threads. AMPL is not, I think, and a large part of the simulation time is in AMPL. So given Frank's comment, that is probably the explanation? The frequency of each core is lower and my simulations are not properly multi-threaded. – evencoil – 2015-12-01T17:46:16.217
All of those are "lower is better" except for CryptoHash. So the new system is much better on Blowfish and CryptoHash, roughly the same on Fibonacci, N-Queens and FFT, and significantly worse on Raytracing. – evencoil – 2015-12-01T17:47:55.190
If its "lower is better" then
CPU Blowfish
andCPU CryptoHash
are clearly multi-threaded given both numbers are about 4x faster which makes sense considering you have 32 additional cores on the new system. The difference withCPU Fibonacci
andFPU FFT
is likely a frequency related and memory related (old system had more memory). The other two are likely closer to want Frank has already said. – Ramhound – 2015-12-01T17:51:07.643Thanks. Minor correction: the new system has a ton more memory (I think you may have dropped a digit). But the explanation makes sense in any case. – evencoil – 2015-12-01T17:55:06.747
@evencoil - nitpicking here....Units are important. If you have 131,957MB of memory your system has 132GB of memory. I realize you were providing us exact ouputs, and appreciate that, but units make it so much easier to read :$ – Ramhound – 2015-12-01T18:05:43.337