1.5Gb memory might be an issue. I wonder if that's going to be the same when we get past the tuning run.
All my 4 core/8 thread i7's only have 8Gb. Sounds like time to upgrade their memory. In the mean time I suppose I can limit the number of concurrent tasks.
I presume CV has avx instructions and G will be a more legacy friendly instruction set.
I think one or both has some degree of provision for detecting and using at least one degree of optional instruction set support.
Both of the machines on which I've received Multi GW work got the G flavor, for which the application name comes up as 1.01 Multi-Directed Continuous Gravitational Wave search Tuning run G (AVX)
I think somewhere an authority here explained that this version does not required AVX, but can use it if available.
One is a newish but pretty low-end i3-2377M, while the other is a much older but higher-end for the time Xeon E5620.
As to memory use, when I've checked the units I got were causing about 650M memory use.
I have also noticed that my 32 bit Windows computer (AMD Phenom 955) does a "G" type in around 18,000 seconds using SSE2 instruction set.
My Linux 64 bit Phenoms take around 38,000 seconds with the AVX in the work unit name, as I don't have AVX I would guess that it is using SSE2 but why is it so much slower than my 32 bit Windows machine?
I have also noticed that my 32 bit Windows computer (AMD Phenom 955) does a "G" type in around 18,000 seconds using SSE2 instruction set.
My Linux 64 bit Phenoms take around 38,000 seconds with the AVX in the work unit name, as I don't have AVX I would guess that it is using SSE2 but why is it so much slower than my 32 bit Windows machine?
It is probably just the variation in the work units rather than the machines. On an i7-4771 (4 cores only) on Win7 64-bit, I get from 23 minutes for the 34 Hz CasA work units to 9 hours for the 923 Hz ones. (Though the 34 Hz ones were during the Beta, and they may be shorter for other reasons. But I would check some more before reaching conclusions as to the machines or OS.)
I have spent a few minutes looking over the Linux hosts which are generating errors and comparing against those which are generating results.
I can not find any Intel AVX enabled processors generating valid GW results on Linux hosts.
I have a i7-860 (does not support AVX) - running nicely the CV tasks with the AVX binary einstein_O1MD1TCV_1.01_x86_64-pc-linux-gnu__AVX - so i guess it has stepped down to SSE2.
I have spent a few minutes looking over the Linux hosts which are generating errors and comparing against those which are generating results.
I can not find any Intel AVX enabled processors generating valid GW results on Linux hosts.
I have a i7-860 (does not support AVX) - running nicely the CV tasks with the AVX binary einstein_O1MD1TCV_1.01_x86_64-pc-linux-gnu__AVX - so i guess it has stepped down to SSE2.
Has anyone seen an AVX enabled processor working with Linux?
I did some preliminary analytics yesterday to check if this is a data or application problem. It is discussed here: https://einsteinathome.org/content/multi-directed-gw-search#comment-150757. What I did is make a list of CPUs that return tasks with exit code 38 and CPUs that return successful results. Both on Linux. I then compared the lists and investigated the models that appeared on both (this list was non empty, which means that it is not strictly CPU model dependent). I'm doing that again with AVX support in mind. So if this indeed is a problem with AVX support, I shouldn't find any successful tasks from AVX enabled CPU's.
Conan_4 wrote:What is the
)
This point is discussed in the first point of this thread.
1.5Gb memory might be an
)
1.5Gb memory might be an issue. I wonder if that's going to be the same when we get past the tuning run.
All my 4 core/8 thread i7's only have 8Gb. Sounds like time to upgrade their memory. In the mean time I suppose I can limit the number of concurrent tasks.
BOINC blog
archae86 wrote:Conan_4
)
Thanks for that but you have to admit it is a bit hidden in the wording, I had to re-read it a few times to work out what they meant.
So a "CV" work unit is Cassiopeia A and Vela Junior work units combined to become CV and sent to fast computers.
A "G" work unit is a G347.3 work unit and is sent to all other computer types.
Conan
Conan_4 wrote:What is the
)
CV is short for CasA and VelaJr which are suppose to go to faster computers
and G is for G347.3 which all computers can get.
I haven't gotten any of the new WUs yet but I presume CV has avx instructions and G will be a more legacy friendly instruction set.
MindCrime wrote:I presume CV
)
I think one or both has some degree of provision for detecting and using at least one degree of optional instruction set support.
Both of the machines on which I've received Multi GW work got the G flavor, for which the application name comes up as
1.01 Multi-Directed Continuous Gravitational Wave search Tuning run G (AVX)I think somewhere an authority here explained that this version does not required AVX, but can use it if available.
One is a newish but pretty low-end i3-2377M, while the other is a much older but higher-end for the time Xeon E5620.
As to memory use, when I've checked the units I got were causing about 650M memory use.
I have also noticed that my
)
I have also noticed that my 32 bit Windows computer (AMD Phenom 955) does a "G" type in around 18,000 seconds using SSE2 instruction set.
My Linux 64 bit Phenoms take around 38,000 seconds with the AVX in the work unit name, as I don't have AVX I would guess that it is using SSE2 but why is it so much slower than my 32 bit Windows machine?
Strange to say the least.
Conan
Conan_4 wrote:I have also
)
It is probably just the variation in the work units rather than the machines. On an i7-4771 (4 cores only) on Win7 64-bit, I get from 23 minutes for the 34 Hz CasA work units to 9 hours for the 923 Hz ones. (Though the 34 Hz ones were during the Beta, and they may be shorter for other reasons. But I would check some more before reaching conclusions as to the machines or OS.)
I have spent a few minutes
)
I have spent a few minutes looking over the Linux hosts which are generating errors and comparing against those which are generating results.
I can not find any Intel AVX enabled processors generating valid GW results on Linux hosts.
I have a i7-860 (does not support AVX) - running nicely the CV tasks with the AVX binary einstein_O1MD1TCV_1.01_x86_64-pc-linux-gnu__AVX - so i guess it has stepped down to SSE2.
see Tasks for host 4918234
Has anyone seen an AVX enabled processor working with Linux?
AgentB wrote:I have spent a
)
I did some preliminary analytics yesterday to check if this is a data or application problem. It is discussed here: https://einsteinathome.org/content/multi-directed-gw-search#comment-150757. What I did is make a list of CPUs that return tasks with exit code 38 and CPUs that return successful results. Both on Linux. I then compared the lists and investigated the models that appeared on both (this list was non empty, which means that it is not strictly CPU model dependent). I'm doing that again with AVX support in mind. So if this indeed is a problem with AVX support, I shouldn't find any successful tasks from AVX enabled CPU's.
Just don't let this be like
)
Just don't let this be like CSG where a user finds a data calculation error and then the application is shut off.