Re: answer // further // Benchmark - 64 bit much
- To: mathgroup at smc.vnet.net
- Subject: [mg111884] Re: answer // further // Benchmark - 64 bit much
- From: "Scot T. Martin" <smartin at seas.harvard.edu>
- Date: Tue, 17 Aug 2010 01:20:34 -0400 (EDT)
Thanks for the several replies on this issue. As there has been this
interest expressed on the correlation between memory hardware and
Benchmark[] (and the rather large surprising dependence), here is the
final report of the different scenarios below. These were all obtained
with the exact same machine but within different memory cards
inserted/withdrawn.
Initial condition, benchmark score of 1.13, obtained using 8 GB but with 2
GB of the "wrong memory" (i.e., a single card plugged into a "tri" memory
slot).
Removal of 2 GB to drop to 6 GB of "correct memory" increased score to
5.50.
Today, I installed the correct memory (from Dell, paid extra, didn't want
headache) to move to 12 GB. The Benchmark[] is now 7.30.
These are the scores in reference to a Dell Preicsion T3500 using 64-bit
Windows 7 on an Intel(R) Xeon(R) CPU 3.33 GHz W3680.
{"MachineName" -> "profmartinoffic",
"System" -> "Microsoft Windows (64-bit)",
"BenchmarkName" -> "MathematicaMark7", "FullVersionNumber" ->
"7.0.0", "Date" -> "August 16, 2010", "BenchmarkResult" -> 7.27,
"TotalTime" -> 11.856, "Results" -> {{"Data Fitting", 0.562},
{"Digits of Pi", 0.328}, {"Discrete Fourier Transform", 0.312},
{"Eigenvalues of a Matrix", 0.827}, {"Elementary Functions",
0.686}, {"Gamma Function", 0.218},
{"Large Integer Multiplication", 0.452}, {"Matrix Arithmetic",
1.622}, {"Matrix Multiplication", 0.749},
{"Matrix Transpose", 0.796}, {"Numerical Integration", 1.622},
{"Polynomial Expansion", 1.076}, {"Random Number Sort", 1.03},
{"Singular Value Decomposition", 0.702},
{"Solving a Linear System", 0.874}}}
On Mon, 16 Aug 2010, telefunkenvf14 wrote:
> On Aug 14, 5:27 am, Joseph Gwinn <joegw... at comcast.net> wrote:
>> In article <i438i1$hs... at smc.vnet.net>,
>> "Scot T. Martin" <smar... at seas.harvard.edu> wrote:
>>
>>> In case any one else runs into a similar problem, I sorted out the answ=
> er
>>> to the slowness below. I pulled out one memory strip to drop from 8 MB =
> to
>>> 6 MB and, presto, performance went from slow (rating of 1.13) to ultraf=
> ast
>>> (rating of 5.50).
>>
>>> How to explain this? I believe it goes back to "following the
>>> manufacturer's recommendation". That is, this computer is provided in
>>> configurations of 6 MB or 12 MB. My IT department bought a 2 MB memory
>>> strip and pumped it up from 6 to 8 MB. It seems there must be some kind=
> of
>>> hardware inbalancing in the memory that influences the speed of some
>>> calculations; details are beyond my technical level. But the problem ha=
> s
>>> been fixed by "following the manufacturer's suggestion on memory
>>> configurations".
>>
>> I can guess what the problem is. Most memory systems on PCs are set up=
> to
>> support a dual-bank architecture, where there is a pair of identical memo=
> ry
>> banks that are accessed in alternation: all the odd-numbered words go =
> in one
>> bank, and all the even-numbered words go in the other. This allows muc=
> h of the
>> memory system overhead to be mitigated.
>>
>> Larger computer systems may use four banks.
>>
>> What is meant by "identical memory banks"? The memory sticks in the
>> corresponding sockets must be the *exact* same type, so one necessarily g=
> ets
>> memory sticks in matched pairs. Most commonly, all memory sticks in a =
> bank must
>> also be identical. There are also other platform-dependent rules on al=
> lowable
>> memory configurations, but not all allowable configurations are also opti=
> mal.
>>
>> Joe Gwinn
>
> I was going to say the same thing. I have triple-channel memory in my
> desktop (3x2GB) with one slot open. My motherboard instructions state
> that any memory added to the last slot will run in 'single channel
> mode' (less bandwidth, likely slower)... I must say though, I would
> *NOT* expect to see differences as dramatic as you report!!!
>
> Check the memory settings in your bios. My guess is that your machine
> is automatically throttling back voltages, clock timings, etc., to ALL
> your dimms, and probably to a VERY low and safe setting. I'm 99%
> positive this is due to your IT department plugging in 'any old
> dimm'---rather than one exactly like the others, or at the very least,
> one with the same specs.
>
> If your bios settings confirm suspicion, here's what you need to do:
> Look at one of the original dimms in your system, find the serial or
> model number and start Googling for the specs. Once you have them, go
> shopping!!! (newegg.com is pretty good; 2GB dimms of DDR3 should be
> cheap) Note: You likely do *NOT* need an exact match---differing model
> numbers and brands may be ok---but the dimm you purchase *needs to
> have* the same CAS latency timings, operating voltages, and type
> (DDR3, I'm guessing).
>
> -RG
>
>