How many flops are contributed to einstein@home?
6 Feb 2010 17:35:17 UTC
Topic 194762
(moderation:
Hi there,
I am searching for numbers on how many flops are contributed to the einstein@home project by the participants.
I am going to hold a talk about einstein@home and want to compare the computing power of various supercomputers with that provided to einstein@home.
Thank you in advance
How many flops are contributed to einstein@home?
)
This is actually not trivial to estimate. You will find some BOINC statistics sites that compute FLOPs numbers for each BOINC project http://de.boincstats.com/stats/project_graph.php?pr=einstein, but the formula for calculating this is rather simplistic and arbitrary.
You can do a bit better by scaling the throughput of E@H as a whole by the time it takes an E@H task to complete on a single core of a supercomputer with known Flops rating and known number of cores (because some idle time of an actual supercomputer, ATLAS at the Albert Einstein Institute in Hannover, Germany, is donated to E@H).
I did this calculation once and IIRC, the performance I got at that time was about 70 TFlops for the E@H network, but that was some time ago. I would not be surprised if it were close to 100 TFlops now.
CU
HBE
This is a helpful idea, thank
)
This is a helpful idea, thank you.
I may just be unable to find information on this topic on your websites, but are there specific numbers on the size of the jobs that are distributed to participants?
I mean something like how many seconds of measuring the inferometers for example, the number of floating point operations contained in each job or an average number of credit returned for each job of e@h.
Another question which should propably precede the first one is if all jobs are of the same size. (I guess so)
If the answers to my questions are documented somewhere I apologize :-)
Hi, I also believe it
)
Hi,
I also believe it would be nice to have a regularily updated figure (once per year?) in FLOPS to compare with current supercomputers of the TOP500 list. My estimation is that BOINC projects are supercheap if we look at the hardware, but maybe needs a little more software engineering efforts because of the BOINC middleware and different client versions. Of course this applies only to very simply to parallel problems, which have no interdependences between workunits.
Regarding the efficiency of the consumed electricity, the current benchmark is at 65 GFLOPS/Watt (Green500). I believe that a BOINC-Cluster is less efficient, because many hosts use not current state of the art hardware and an optimised system design which you would choose for a specific number crunching task like simulating nuclear fusion and fission, meteorology, etc.
Is there an upper threshold for the electrical energy efficiency, if we assume that the first and second law of thermodynamic also applies to converting data in some kind of order with a lower entropy?
Regards,
Gunnar