Windows XP-SP2 32bits VS Windows 7-SP1 64bits
10 Nov 2013 15:38:03 UTC
Topic 197263
(moderation:
Hello,
I run EaH on a Windows XP-SP2 (32bits) machine for several years and recently I install (dual-boot) Windows 7 SP1 (64bits).
I notice a large different in the boinc benchmarks (especially in the floating point MIPS) between the two OSs:
WinXPSP2(32bits) [Boinc 7.0.27]-> 3440/8050 (floating point/integer MIPS)
Win7SP1 (64bits) [Boinc 7.2.28]-> 2806/7513 (floating point/integer MIPS).
The PC has an AMD PhenomII x4 965 with 4GB DDR2-800.
I also notice that in the Win7 (64bits) the EaH client is 32bit.
Is there any 64bit EaH client?
Is there any comparison between XP 32bits and Win7 64bits for the EaH clients?
Thank you,
Windows XP-SP2 32bits VS Windows 7-SP1 64bits
)
OK, related:
Is this a myth or for real?
On current 64-bit hardware, in general, Windows 32-bit is actually faster than Windows 64-bit!
Really? How? Why?
???
Happy fast crunchin',
Martin
See new freedom: Mageia Linux
Take a look for yourself: Linux Format
The Future is what We all make IT (GPLv3)
RE: ... I notice a large
)
I wouldn't get too concerned about BOINC benchmarks. My impression is that they can vary quite considerably with different versions of BOINC on the same hardware.
The full range of available science applications for all the different platforms can be inspected through the front page link on the web site. To my knowledge they are essentially all 32-bit apps, although some of those listed are designed to run on a 64-bit version of the OS. As stated on that page, the correct app for your platform will be sent to you automatically. I recall various postings from the Devs over the years indicating that there wouldn't be any real advantage to be gained from a 64-bit science app - or something along those lines - it was quite a while ago :-).
I believe there will be if you publish your 'before' and 'after' results :-). I wouldn't expect any significant difference though.
Cheers,
Gary.
RE: Is this a myth or for
)
Mostly it is true, but we're talking about 0 - 3% differences here. It arises from software which can't take advantage of the 64 bit benefits (which is actually the majority, otherwise we'd have transitioned much earlier) but gets hampered a bit by the drawbacks (larger code means smaller effective cache sizes and more bandwidth needed for transfers).
But then there is some 64 bit software which does profit.. here performance gains can vary greatly and by far exceed 10%. And there's the ability to use more memory, of course.
MrS
Scanning for our furry friends since Jan 2002
So, after running some S6 WUs
)
So, after running some S6 WUs (CasA-SSE2) in Win7SP1 64bit I didn't see any performance difference. In fact WinXPSP2-32bits was 1.5% more efficient with CasA-SEE2 client !!!
These results came from an AMD PhenomII x4 965 @ 3.4GHz running 2 WUs in parallel.
I have another PC with an AMD PhenomII 1090T x6 @ 3.4GHz runing 3 WUs in parallel with a gentoo linux amd64.
In linux there is a CasA 64bit client and the results from this PC is about 33% more efficient.
Keep in mind that the 2 processors are in the same frequency, 45nm SOI with L3 cache of 6MB both and that the X6 has 1MB more L2 cache. Furthermore, the PC with X6 have DDR2-1066 while the PC with X4 have DDR2-800.
I don't think that 266MHz difference in the DDR2 and 1MB more L2 cahce will produce 33% more efficient results in CasA client.
I can't know the performance gain with linux vs windows clients but I assume that we can see a big gain with 64bits EaH clients in windows...
How do you measure
)
How do you measure efficiency?
And E@H uses floating point maths, so it's either single precision (32 bit) or double precision (64 bit, internally 80 bit before rounding) independent of the OS being 32 or 64 bit. The only advantage to compiling a 64 bit app would be the enhanced number of registers available.
I'm not an insider, but I assume the project team has already looked into this option, considering the massive amount of crunching needed by E@H.
BTW: without knowing your measurment methods I'd say a performance difference of 1.5% is nothing, or at best borderline to the noise level. Which is indeed the expected performance difference between different Win versions for regular CPU crunching.
MrS
Scanning for our furry friends since Jan 2002
I think we only use single
)
I think we only use single precision float representations for our WU's at E@H .... ???
As you say, that's a distinct question from the general word/address length for a given OS.
Cheers, MIke
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
Any WUs also done on various
)
Any WUs also done on various GPUs are single precision, for sure. But I'm not sure about the CPU-only WUs. One reason we're not seeing GPU versions of them yet might be that they require double precision. That's just my speculation, though.
MrS
Scanning for our furry friends since Jan 2002
RE: How do you measure
)
Hi,
Efficiency is the % difference in processing time of the same WUs. There is 1.5% difference in processing time (CasA-SSE2) of the 32bit EaH client in WindowsXP 32bit and Windows7 64bit.
Comparing the 32bit CasA client (windows7) with the 64bit CasA client (Linux) I found a 33% difference, running on different PCs with similar CPUs @ 3.4GHz both.
From statistics on Gamma-ray pulsar search client which is 32bit only (both in windows and linux) I found that linux client is ~19% more efficient (less processing time) than the windows client.
So we have the following statistics:
PC1 (Windows7-64bits) :
a) Gamma-ray (32bit client) -> ~22h / WU
b) CasA (32bit client) -> ~9.5h / WU
PC2 (Linux 64bits) :
a) Gamma-ray (32bit client) -> 17.8h / WU ( ~19%)
b) CasA (64bit client) -> ~6.5h / WU ( ~32%)
If we assume (for simplicity) that the 19% difference in Gamma-ray client times is due to differences between the two PCs (OS, memory, mb, etc) then there is a ~13% "clear" difference between CasA 32bit and 64bit clients.
OK, so your efficiency is
)
OK, so your efficiency is "performance per CPU clock".
Anyway, I'm not convinced your conclusion
can be drawn from this data. The reason: different compilers are being used / necessary for the different platforms. You never know what these may change, depending on OS and 32/64 bit.
There's also the issue of E@H being very sensitive to main memory bandwidth and latency. Both factors favor your Linux system.
Yet I completely agree with the following: based on these observations there is very probably quite some potential for optimization left in the current Win client - be it due to 64 bit or compiler optimizations.
MrS
Scanning for our furry friends since Jan 2002