Message boards :
Number crunching :
How do I consolidate a bunch of systems to they "look" like one system to BOINC?
Message board moderation
Author | Message |
---|---|
![]() Send message Joined: 28 Nov 02 Posts: 5126 Credit: 276,046,078 RAC: 462 ![]() |
I wonder if the tools are readily available to allow say something like a Beowulf cluster to process the tasks but from the "outside world" point of view it is a single client. The Supercomputer Clusters are clearly showing a "unified" POV but have a lot of hardware and processing spread all over the place. Can we do this with consumer hardware? Or consolidate a PI cluster so it looks like a single computer to BOINC? Tom A proud member of the OFA (Old Farts Association). |
rob smith ![]() ![]() ![]() Send message Joined: 7 Mar 03 Posts: 22657 Credit: 416,307,556 RAC: 380 ![]() ![]() |
You need to do a bit of reading - there are quite a few people who have built Beowulf clusters using RPi, and some are better documented than others. Your idea should work, but be aware that BOINC may not be happy seeing so many cores.... I'm sure that there was someone running one in the past - do a search of these boards. Bob Smith Member of Seti PIPPS (Pluto is a Planet Protest Society) Somewhere in the (un)known Universe? |
![]() Send message Joined: 28 Nov 02 Posts: 5126 Credit: 276,046,078 RAC: 462 ![]() |
You need to do a bit of reading - there are quite a few people who have built Beowulf clusters using RPi, and some are better documented than others. Thank you. I sounds interesting and I am curious. :) I will post links as I run across them. This almost certainly is going to take a while to see if I can even get to a starting position. Tom A proud member of the OFA (Old Farts Association). |
rob smith ![]() ![]() ![]() Send message Joined: 7 Mar 03 Posts: 22657 Credit: 416,307,556 RAC: 380 ![]() ![]() |
One thing to consider when thinking about running SETI on a Beowulf cluster - it will appear to the servers as a single multi-core CPU, not a CPU plus a number of "GPU", thus will be restricted in the number of tasks it will get from the server. Edit: Just one other thing, the SETI applications are single threaded (or at least written to use one core per instance), so you will still be stuck with the multi-hour/task execution times :-( Bob Smith Member of Seti PIPPS (Pluto is a Planet Protest Society) Somewhere in the (un)known Universe? |
Cygnus X-1 Send message Joined: 15 Feb 04 Posts: 75 Credit: 3,732,505 RAC: 175 ![]() ![]() |
There is an application called "BoincTasks" that allows you to view multiple computers in a single Boinc Manager like window. It doesn't look like a single system to Boinc but it allows you to monitor them as if they were. https://efmer.com/ |
rob smith ![]() ![]() ![]() Send message Joined: 7 Mar 03 Posts: 22657 Credit: 416,307,556 RAC: 380 ![]() ![]() |
Efmer's Boinc Tasks is nothing like a Beowulf Cluster. A Beowulf Cluster is a cluster of similar (or identical) "small" computers that act as one "large" one, whereas Boinc Tasks allows one to manage a lot of computers all acting independently of each other from a single computer. Edit: If one is only running SETI on a pile of RPi then something like BoincTasks is going to be better than a Beowulf cluster. Bob Smith Member of Seti PIPPS (Pluto is a Planet Protest Society) Somewhere in the (un)known Universe? |
![]() Send message Joined: 28 Nov 02 Posts: 5126 Credit: 276,046,078 RAC: 462 ![]() |
https://www.zdnet.com/article/build-your-own-supercomputer-out-of-raspberry-pi-boards/ A proud member of the OFA (Old Farts Association). |
Ville Saari ![]() Send message Joined: 30 Nov 00 Posts: 1158 Credit: 49,177,052 RAC: 82,530 ![]() ![]() |
Modify boinc client to see more cpu cores than the host it is running in really has and write a 'proxy' setiathome application that runs the real setiathome application in a remote host. Or I guess this might work even without modifying boinc client by making the app_info.xml claim the proxy app requires less than 1.0 cpus. |
rob smith ![]() ![]() ![]() Send message Joined: 7 Mar 03 Posts: 22657 Credit: 416,307,556 RAC: 380 ![]() ![]() |
Neither will work - it is a hard, server-side limit, that restricts users to one-CPU-worth (150?) tasks even on mulit-CPU systems, and the CPU applications are written to use all of each available CPU core, not fractions or multiples thereof. Bob Smith Member of Seti PIPPS (Pluto is a Planet Protest Society) Somewhere in the (un)known Universe? |
![]() Send message Joined: 28 Nov 02 Posts: 5126 Credit: 276,046,078 RAC: 462 ![]() |
Basically, in order to make a "farm" more competitive on the Leaderboard, I was looking at some way to have them crunch under a single BOINC/Seti@Home id. I acknowledge the specific limitations on the CPU tasks would make this problematic for high core counts. There is at least one EPYC box here that has turned off his SMT because he couldn't use all 128 threads. Tom A proud member of the OFA (Old Farts Association). |
rob smith ![]() ![]() ![]() Send message Joined: 7 Mar 03 Posts: 22657 Credit: 416,307,556 RAC: 380 ![]() ![]() |
Before you get too carried away with a big pile of RPi do a quick sum to work out how many you would need. A single RPi has four cores, each core takes about 8 hours to do a single task, a fairly recent i7 core does a single task in what? (Let's say an hour) So we need 8 times as many RPi cores as i7 cores to do the same work per hour, which is two complete RPi to do the work of one i7 core, so that is 16 RPi would give roughly the same output as an i7 on its own. I'm not sure how far up the tree that is, but I do know it's a long way from being top 100 :-( Bob Smith Member of Seti PIPPS (Pluto is a Planet Protest Society) Somewhere in the (un)known Universe? |
![]() ![]() ![]() ![]() Send message Joined: 1 Dec 99 Posts: 2786 Credit: 685,657,289 RAC: 835 ![]() ![]() |
With the growth of threads in the current CPUs it might be better off seeing if the developers are interested in programming a multithread (4?) CPU app. ![]() |
![]() Send message Joined: 28 Nov 02 Posts: 5126 Credit: 276,046,078 RAC: 462 ![]() |
Before you get too carried away with a big pile of RPi do a quick sum to work out how many you would need. A single RPi has four cores, each core takes about 8 hours to do a single task, a fairly recent i7 core does a single task in what? (Let's say an hour) So we need 8 times as many RPi cores as i7 cores to do the same work per hour, which is two complete RPi to do the work of one i7 core, so that is 16 RPi would give roughly the same output as an i7 on its own. +1 Thank you for the analysis. It looks like some kind of Network of say Amd 2700x's would make more sense after than analysis. Back to the future :) Tom A proud member of the OFA (Old Farts Association). |
![]() Send message Joined: 28 Nov 02 Posts: 5126 Credit: 276,046,078 RAC: 462 ![]() |
With the growth of threads in the current CPUs it might be better off seeing if the developers are interested in programming a multithread (4?) CPU app. An interesting question. I know that on my Amd 3950x when I was running it with an AVX cpu task it was working harder and only a little faster. It seems like we would need a consumer-level example of a 3-4 thread per core cpu before it would make sense to try developing something. I suspect that a "gpu server" attack may make more sense. Tom A proud member of the OFA (Old Farts Association). |
rob smith ![]() ![]() ![]() Send message Joined: 7 Mar 03 Posts: 22657 Credit: 416,307,556 RAC: 380 ![]() ![]() |
The n-core approach has been discussed, but there are a few things that need to be considered. Not the least of which is the synchronization of the outputs from the cores. GPUs are designed with this in mind, but CPUs are not quite so good, and the gains in performance are not exactly linear in my experience (I frequently have to manipulate large 3d AutoCad models and the gain in performance using the same model when restricted to use two cores vs eight only affects the re-ceate time by a factor of two, not the expected factor of four.) Bob Smith Member of Seti PIPPS (Pluto is a Planet Protest Society) Somewhere in the (un)known Universe? |
HighTech67 Send message Joined: 22 Jun 03 Posts: 20 Credit: 5,936,475 RAC: 126 ![]() ![]() |
The n-core approach has been discussed, but there are a few things that need to be considered. Not the least of which is the synchronization of the outputs from the cores. GPUs are designed with this in mind, but CPUs are not quite so good, and the gains in performance are not exactly linear in my experience (I frequently have to manipulate large 3d AutoCad models and the gain in performance using the same model when restricted to use two cores vs eight only affects the re-ceate time by a factor of two, not the expected factor of four.) This non-linear scaling is very true. Ask anyone over at PrimeGrid. Their LLR tasks are MT (multi-thread) capable and you don't cut single-core run time by the number of cores used to multi-thread. It is some percentage that depends on the machine in question, the number of cores actually used, and the type of task as to the gain you achieve. |
Ville Saari ![]() Send message Joined: 30 Nov 00 Posts: 1158 Credit: 49,177,052 RAC: 82,530 ![]() ![]() |
Before you get too carried away with a big pile of RPi do a quick sum to work out how many you would need. A single RPi has four cores, each core takes about 8 hours to do a single taskHow much electrical power are modern RÏ€ consuming? It one completes 4 tasks in 8 hours, then you need about 80 RÏ€ to match the output of my single 130W GPU. |
Ville Saari ![]() Send message Joined: 30 Nov 00 Posts: 1158 Credit: 49,177,052 RAC: 82,530 ![]() ![]() |
The n-core approach has been discussedPerformance gains from multithreading can never exceed linear and and they can match that only in very special cases. But when you run multiple independent tasks in parallel, you get linear gains by default if the core performance is the bottleneck instead of disk or memory bandwidth. So there's no reason to even try to develop a multithreaded seti app because the best performance you could theoretically reach is no better than the performance of multiple singlethreaded apps running in parallel. |
rob smith ![]() ![]() ![]() Send message Joined: 7 Mar 03 Posts: 22657 Credit: 416,307,556 RAC: 380 ![]() ![]() |
My RPi4 are drawing between 8 and10 watts from the wall, each has four cores. Let's assume we have an 8-core processor, with all its support stuff in a PC drawing 130 watts from the wall. Using the figure I gave earlier we would need 16 RPi4 to deliver the same throughput, so the power draw would be between 128 and 160 watts from the wall, somewhere between much the same and 25% more. The reason I gave two figures is that some of my RPi4 are attached to a single PSU which is more efficient that the individual PSUs on the others, until I looked at the figures a few minutes ago I didn't realise just how much more! Bob Smith Member of Seti PIPPS (Pluto is a Planet Protest Society) Somewhere in the (un)known Universe? |
![]() Send message Joined: 28 Nov 02 Posts: 5126 Credit: 276,046,078 RAC: 462 ![]() |
So if I just go out and get an EYPC cpu/mb, a lot more gpus and an entire power system for the house I can "look" like one system :) A proud member of the OFA (Old Farts Association). |
©2025 University of California
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.