Has anyone built a beowulf cluster?

[FindersKeepers]

 .. or worked on one? I am interested in knowing how practical they are, what real world applications they have been used for in engineering, construction and manufacture. Datamining I can understand for medical/healthcare but would like to know if someone here has experience in or around one.

Some cheap fast hardware out is in the second hand market now. I have components to start playing with my own.

[Dingo]

I used to have nine computers that were linked a long time ago and I also setup a number of diskless PC's using DRBL.     http://drbl.sourceforge.net/

The idea is that you only have one machine with the hard drive and all the other machines just use their processing power to crunch.    Good for teaching and such as that was one of the reasons I was looking at it as a closed network.   It ended up being slower for BOINC than normal operations and disks are not expensive but it was something to play with for a while.

[FindersKeepers]

Thanks Dingo, will have a look. I was thinking distributed computing to partition each processor with it's own local disk storage using various static source data for datamining. 1 processor (1 thread) searching 20 gig is bottlenecked to billioh but 20 processors each searching 1gig is a whole lot faster. I would imagine the diskless remote thing would be throttled by the gigabit ethernet when using large data sets. Making use of fast threaded i5 4 processor nodes with a 6G/s HDD and 2 channel memory would then be a memory bottleneck.  I have tested it scaling 8 processors distributing 8 gig data on a local machine with 2 local hard drives. But datamining/analysis is quite simple - apply the same to databases is similar.

Would only need small local HDD's. Using PCIe M2 SSD's would be a little bit awesome but probably overkill. If a node fails can replace it from a master store.

[FindersKeepers]

*bump* .. wondering if this project is worth pursuing?

Linux cluster distro's I have come across ...
Mandriva iggi cluster
Scyld (on redhat)

[fubared]

Not worth the effort these days imho. Clustered computers were great when cpus are single core, now 8 cores are affordable and 10+ is a small premium.

Worthless for boinc as its simplier just to have a bunch of independent systems controlled by one manager.

[Mike Mitchell]

I would agree with that.

[FindersKeepers]

Thanks - I agree on boinc. I understand this is a boinc forum but was considering more distributed/cluster oriented projects. Boinc could benefit by having a local task server that is faster than the internet so that the cpu time spent connected to a remote server is reduced. But that is really penny savings compared to other o/s and platform gains.

[FindersKeepers]

Hi - clarifying why this might become a thing again ... bottlenecks have moved back to the network making nodes faster. It is how boinc is effective on standalone machines. With the advent of fast multichannel memory, 6G/s sata drives (12G now) then scalable node computing is still a powerful tool. Am wondering if someone knows of worthwhile projects. A distributed compiler for instance ... distributed data mining? (google would use one).

All good if nothing applies. Just putting it out there.

[fubared]

If you're making a render farm, cluster computers might work. But these days, unless you are making your own blockbuster, its simpler to get a dual socket EPYC or Xeon, slap on a bunch of video cards and call it a day.

If you are thinking along the lines of supercomputer, you have to make sure every node (computer in the cluster) is more or less identical. From playing with RAID, which is a bunch of drives stitched into a single array, the array will be as fast as the slowest drive. Also, the network. Its not just about bandwidth, its also latency. More computers=more chance of packet clashing=higher latency. Elephant in the room is probably the software. Right now software can't really handle Ryzen's CCX clusters, imagine trying to navigate multi CPUs on different motherboards with differing latency. You might have to compile your own. This limits you to open source or home brew.

But if you have spare time try it and report back, some of us are interested.

[FindersKeepers]

One very large positive to come out of playing with Linux ... have had to standardised the encoding of all my data files to Linux encoding to get dos2unix to work properly. The resulting reduction in raw data file size is 45% - 9gig in data went down to 5gig. Even windoze has benefited from it.

Unintended consequences I guess. 

[FindersKeepers]

Data retrieval on a 1Gig data set is an order of magnitude (10x) faster on Linux. Insane!!
Going to be a little bit of fun playing with a cluster of these things.

Giddyup.