Einstein and x86_64
5 Mar 2005 11:13:51 UTC
Topic 188276
(moderation:
I am running Einstein at home on two Linux computers. One is an AMD Barton core 2500+ (1.87 GHz, 32bit) and the second is an AMD64 2600+ (1.9GHz, 64bit). The OS is Fedora Core 3 of the appropriate bit-ness. The Einstein client on the 32 bit computer finishes a work unit in 42k seconds, while the 64 bit processor takes 78k CPU seconds. These times are consistent over 5 work units on each computer. That seems a remakable penalty for the 64 bit computer, especially if Einstein at home uses double precision floating point calculations. The 64 bit computer should be twice as fast as the 32 bit, not vice versa.
Einstein and x86_64
)
> The 64 bit computer should be twice as fast as the 32 bit, not vice versa.
I can't comment on this other than to commiserate with you on this woeful performance. Another user has mentioned that running the Windows client under WINE on linux has given a dramatic increase in performance. Sorry I don't have a link but do a search on WINE and you are bound to get all the discussion.
Good luck!! Your comps deserve much better than those times.
Cheers,
Gary.
I'm using an Athlon 64 3000+
)
I'm using an Athlon 64 3000+ 2.0 GHz overclocked to 2.2GHz and it's finishing at 20k seconds, your 64 should be much faster than what you seem to be getting, my OS is Windows XP Pro
My Athlon 3000+ (not
)
My Athlon 3000+ (not overclocked) on Linux takes 37k seconds per.
---------------------------------------------------------------------------
stlforsale.com
> The 64 bit computer should
)
> The 64 bit computer should be twice as fast as the 32 bit, not vice versa.
If and only if the processor architectures are identical, then in theory, yeah.
If the FPU on both parts are the same width (80 bits for IEEE standard units) and you're doing LOTS of floating point, probably not.
I've an Athlon64 3500+ @
)
I've an Athlon64 3500+ @ 2.475GHz. Windows XP Pro x64. Currently have work units coming in at 18K sec.
Much looking forward to the appearance of 64-bit versions of BOINC and Einstein. . .
The 64 bit processor requires
)
The 64 bit processor requires a 64 bit OS and the app has to be compiled to take advantage of the 64 bit bandwidth. Without this the Athlon 64 is running in 32 bit emulation mode.
I'm running an Athlon 2700 and it is turning out WU in 20.5k sec. This appears to be the price performance break point for the Athlon with the current OSes available.
> The 64 bit processor
)
> The 64 bit processor requires a 64 bit OS and the app has to be compiled ...
Indeed, when I compiled SETI@Home on my machines it seem to run about 40% faster. That is not yet an option with Einstein I guess.
---------------------------------------------------------------------------
stlforsale.com
My understanding was that
)
My understanding was that Einstein@Home was doing a lot of double precision floating point calculations. These would require two memory accesses for each number in 32 bit, but only one in 64 bit mode. Thus my comment that the Athlon 64 bit would run twice as fast. I believe this is indeed the difference between the performance of Mathematica on a Pentium vs. an AMD64 or G5. This would clearly require a 64bit OS, which I am running.
On a related topic, it seems clear that even the 32 bit linux performance is not nearly that of windows. I like the einstein project, but if they are not going to put effort into optimizing my client, then they are wasting the CPU cycles I'm willing to donate. In that case, it is time for me to find a project that takes more care.
duke out
(ref; http://arstechnica.com/articles/paedia/cpu/amd-hammer-1.ars)
> > The 64 bit computer should be twice as fast as the 32 bit, not vice
> versa.
>
> If and only if the processor architectures are identical, then in theory,
> yeah.
>
> If the FPU on both parts are the same width (80 bits for IEEE standard units)
> and you're doing LOTS of floating point, probably not.
>
> My understanding was that
)
> My understanding was that Einstein@Home was doing a lot of double precision
> floating point calculations. These would require two memory accesses for each
> number in 32 bit, but only one in 64 bit mode. Thus my comment that the
> Athlon 64 bit would run twice as fast. I believe this is indeed the
> difference between the performance of Mathematica on a Pentium vs. an AMD64 or
> G5. This would clearly require a 64bit OS, which I am running.
>
> On a related topic, it seems clear that even the 32 bit linux performance is
> not nearly that of windows. I like the einstein project, but if they are not
> going to put effort into optimizing my client, then they are wasting the CPU
> cycles I'm willing to donate. In that case, it is time for me to find a
> project that takes more care.
>
> duke out
>
> (ref; http://arstechnica.com/articles/paedia/cpu/amd-hammer-1.ars)
>
>
> > > The 64 bit computer should be twice as fast as the 32 bit, not vice
> > versa.
> >
> > If and only if the processor architectures are identical, then in
> theory,
> > yeah.
> >
> > If the FPU on both parts are the same width (80 bits for IEEE standard
> units)
> > and you're doing LOTS of floating point, probably not.
> >
>
>
> > My understanding was that
)
> > My understanding was that Einstein@Home was doing a lot of double
> precision
> > floating point calculations. These would require two memory accesses for
> each
> > number in 32 bit, but only one in 64 bit mode. Thus my comment that the
> > Athlon 64 bit would run twice as fast.
There are so many factors that isolating one and saying "doubling the bus width will double speed" just doesn't quite work.
Besides, the typical modern 32 bit processor already has a 64 bit memory bus for exactly the reason you stated: you get twice as much data during each memory cycle.