Thanks :)
Guess I'm near the top in the failure rate rank list ;)
Other failures happened when trying Fedora C8 and Open Suse 10.3 or switching from a user to root. Downgrading from Boinc 5.10.xx to 5.8.xx doesn't seem to work either.
+some plain clumsiness
Thanks :)
Guess I'm near the top in the failure rate rank list ;)
Other failures happened when trying Fedora C8 and Open Suse 10.3 or switching from a user to root. Downgrading from Boinc 5.10.xx to 5.8.xx doesn't seem to work either.
+some plain clumsiness
Which problems did you get with opensuse10.3 or root-account? i'm using 10.2 now without troubles as root in usermode not damon.
From what I've read on the BOINC developer list, Bernd and David Anderson have identified and essentially fixed the problem in the BOINC source code already this weekend, so the next build of the Linux app will be ok as it will use the corrected BOINC library. So I guess it's not worth the hassle to downgrade BOINC client radically because you will want to upgrade again once the new version of the E@H app is distributed. Not sure how fast the other affected projects can/will react, tho. Note that downgrading the BOINC version will only decrease the propability of the signal 11 issue, the root problem is in the BOINC library that gets linked to the science app (see Bernd's post above).
But, by moment, I have solved the issue using a local DNS server, as I have explained before in this thread. No need to downgrade to an old version of BOINC anymore.
After experimenting a little with it, I can say now that this method is very effective. No matter if the connection fails, all my workunits are finished OK and successfully validated.
It looks like the "signal 11" problem has finally been found and fixed in the new Beta Test App.
Many many thanks to Bikeman and Kathryn, and everyone who helped with reports!
The next "power App" will include this fix, probably some more speedup, and should work on all CPUs that feature SSE (no SSE2 code). I hope to finish it this week, but m not completely sure.
I see a speedup by 20-25% on Intel Core2 boxen but only 10% on AMD K8.
Wonder why?
The main improvement in this version is that it uses SIMD (single instruction multiple data) instructions from the SSE instruction set that can operate in parallel on several floating point numbers. This is the SSE optimization or vectorization that you will read frequently about in this forum, mainly in connection with Akosf :-).
For E@H, this means processing 4 single precision floating point numbers in parallel, packed into a 128 bit (4*32) SSE register.
Now, all the older CPUs from both Intel and AMD (Pentium III, Pentium 4, Pentium M, Core and all AMD K8 CPUs, even the pre-Barcelona Opterons) had only 64 bit wide data paths inside the CPU to feed the SIMD unit, so even a single SSE instruction on 128 bit data would be broken down internally into two 64 bit "micro-instructions".
Basically the SSE implementations in Intel and AMD CPUs were only half as powerful as they could have been.
The Intel Core 2 CPUs (among many other improvements) now finally have a SIMD unit that can swallow 128 bit data in one piece. And if the code is arranged in a nice way, the CPU can even work on several SIMD instructions in parallel.
That's why both the K8 and Intel's pre-Core2 CPUs will not benefit as much from SSE as the Core 2.
In theory the K10 AMD CPUs should also show a huge benefit from SSE code as they also feature similar improvements in the SIMD department, but users reported disappointing results with the new K10 CPUs.
RE: Another 13 units lost
)
try Augustines 5.10.30. so far i had no troubles with network-access.
Thanks :) Guess I'm near the
)
Thanks :)
Guess I'm near the top in the failure rate rank list ;)
Other failures happened when trying Fedora C8 and Open Suse 10.3 or switching from a user to root. Downgrading from Boinc 5.10.xx to 5.8.xx doesn't seem to work either.
+some plain clumsiness
RE: Thanks :) Guess I'm
)
Which problems did you get with opensuse10.3 or root-account? i'm using 10.2 now without troubles as root in usermode not damon.
Hi ho! From what I've read
)
Hi ho!
From what I've read on the BOINC developer list, Bernd and David Anderson have identified and essentially fixed the problem in the BOINC source code already this weekend, so the next build of the Linux app will be ok as it will use the corrected BOINC library. So I guess it's not worth the hassle to downgrade BOINC client radically because you will want to upgrade again once the new version of the E@H app is distributed. Not sure how fast the other affected projects can/will react, tho. Note that downgrading the BOINC version will only decrease the propability of the signal 11 issue, the root problem is in the BOINC library that gets linked to the science app (see Bernd's post above).
CU
Bikeman
Ok, I will wait for the next
)
Ok, I will wait for the next Linux app.
But, by moment, I have solved the issue using a local DNS server, as I have explained before in this thread. No need to downgrade to an old version of BOINC anymore.
After experimenting a little with it, I can say now that this method is very effective. No matter if the connection fails, all my workunits are finished OK and successfully validated.
It looks like the "signal 11"
)
It looks like the "signal 11" problem has finally been found and fixed in the new Beta Test App.
Many many thanks to Bikeman and Kathryn, and everyone who helped with reports!
The next "power App" will include this fix, probably some more speedup, and should work on all CPUs that feature SSE (no SSE2 code). I hope to finish it this week, but m not completely sure.
BM
BM
I see a speedup by 20-25% on
)
I see a speedup by 20-25% on Intel Core2 boxen but only 10% on AMD K8.
Wonder why?
RE: I see a speedup by
)
The main improvement in this version is that it uses SIMD (single instruction multiple data) instructions from the SSE instruction set that can operate in parallel on several floating point numbers. This is the SSE optimization or vectorization that you will read frequently about in this forum, mainly in connection with Akosf :-).
For E@H, this means processing 4 single precision floating point numbers in parallel, packed into a 128 bit (4*32) SSE register.
Now, all the older CPUs from both Intel and AMD (Pentium III, Pentium 4, Pentium M, Core and all AMD K8 CPUs, even the pre-Barcelona Opterons) had only 64 bit wide data paths inside the CPU to feed the SIMD unit, so even a single SSE instruction on 128 bit data would be broken down internally into two 64 bit "micro-instructions".
Basically the SSE implementations in Intel and AMD CPUs were only half as powerful as they could have been.
The Intel Core 2 CPUs (among many other improvements) now finally have a SIMD unit that can swallow 128 bit data in one piece. And if the code is arranged in a nice way, the CPU can even work on several SIMD instructions in parallel.
That's why both the K8 and Intel's pre-Core2 CPUs will not benefit as much from SSE as the Core 2.
In theory the K10 AMD CPUs should also show a huge benefit from SSE code as they also feature similar improvements in the SIMD department, but users reported disappointing results with the new K10 CPUs.
CU
Bikeman
A new "Power App" 4.27 is
)
A new "Power App" 4.27 is available. It should only require SSE and have the 'signal 11' bug fixed.
BM
BM