Parallelism - Cluster Builds, etc.

J

jaymax

Hello,

Could users recommend or suggest any Forum, User groups, books etc. where one can learn about building, architecture, programming etc of clusters / massively parallel computing systems using the FreeBSD O/S and with multiple type nodes. Systems based on GPGPU's enabling desktop facilitation of teraflop capacity are particularly attractive.​

Thanks!
 
General HPC: I don't know of any forum or user group that is general. For specific things, like certain pieces of software (MPI-IO, Lustre, GPFS) there are discussion forums; for commercial software they probably require admission checking.

Books: A google search would find several here; one that comes to mind (because a friend worked on it a few years ago) is "High Performance Parallel IO", edited by Prabhat. Conferences: Look at the proceedings of the "Supercomputing" conference, also known as SC2017, SC2016, and so on. Find those on the web, look at the program of tutorials and technical sessions, and then google the authors and companies involved.

Based on GPUs: Many (perhaps most) current supercomputer builds use GPUs. Whether to use GPUs or not is an interesting question, which is debated in the community.

Using FreeBSD: That's probably hard. I have not heard of any cluster or supercomputer that uses FreeBSD. The last few "top500" lists (which is a list of the 500 largest supercomputers that are non-classified) had Linux at 100% share: every single one uses Linux!

Multiple type nodes: While quite a few compute clusters are made up from different type of nodes, the nodes tend to be homogeneous within a cluster. A typical setup might be: subcluster 1 with 1000 homogeneous machines, bought in 2014 plus 50 IO servers; subclusters 2 and 3 with 600 machines each (one with more memory, the other with GPUs), bought in 2015, subcluster 3 with 1200 more modern machines bought in 2016 and so on. It is really hard to do cluster computing on machines with wildly different performance; even with intelligent workload management, the whole cluster sometimes becomes as slow as the slowest member (the convoy effect).
 
Datapanic said:
There are some, interesting article here: Building a High-performance Computing Cluster Using FreeBSD
Click to expand...
Interesting for whom, historians of scientific computing? FreeBSD version refereed in that article is 4.xxx. The guy who wrote the article was a serious FreeBSD kernel contributor. Most people can't effort to time to become kernel contributors before using clusters to do the science. Unfortunately scientific computing requires some vendor support. I am not taking Juniper networks taking FreeBSD tweaking it and turning into JunoOS. I am talking proprietary compilers, Intel MKL, NVidia proprietary GPU drivers and CUDA framework. FreeBSD has no vendor support to speak of. For better or worse Linux is the only player in that arena. And for the record I can't stand Linux and would pick FreeBSD, with all its flaws, to work with any time day or night over the Linux.
 
In June 2003, 68.9% of the Top500 supercomputers ran "Unix" (meaning a commercial Unix), 27.9% ran Linux, 11.8% ran *BSD, and 0.2% ran Windows. If you break that 68.9% of commercial Unix down further, the biggest players were HP-UX at 24.6%, AIX at 20.4%, Irix at 7.6%, Unicos at 5.2%, Tru64 at 4.2%, Super-UX at 3.0%, UXP/V at 2.2%, HI-UX/MPP at 2.0%, and finally Solaris at 2.0%. The list by performance (not number of systems) is slightly different, with AIX beating every other OS (even Linux), but on that list Solaris doesn't do much better either, at 2.5% market share.

Much has changed since then; today Linux has 100% market share among the supercomputers on the Top500 list. Now that list is not exhaustive; there are smaller supercomputers that don't make the list: the smallest on the list has 17K cores, which probably means less than 1000 nodes, and there are a lot of clusters with dozens or hundreds of nodes in the world. The list also doesn't include some of the largest computers in the world, whose existence is not officially known, and their exact size is classified. But again, I'm pretty sure none of those secret government supercomputers run FreeBSD (although I suspect that some run commercial or proprietary OSes).
 
  • Thanks
Reactions: Oko
ralphbsz said:
In June 2003, 68.9% of the Top500 supercomputers ran "Unix" (meaning a commercial Unix), 27.9% ran Linux, 11.8% ran *BSD, and 0.2% ran Windows. If you break that 68.9% of commercial Unix down further, the biggest players were HP-UX at 24.6%, AIX at 20.4%, Irix at 7.6%, Unicos at 5.2%, Tru64 at 4.2%, Super-UX at 3.0%, UXP/V at 2.2%, HI-UX/MPP at 2.0%, and finally Solaris at 2.0%. The list by performance (not number of systems) is slightly different, with AIX beating every other OS (even Linux), but on that list Solaris doesn't do much better either, at 2.5% market share.
Click to expand...
Thanks for correcting me. I fixed the post. In academia where I worked around that time was all Solaris. I didn't have access to the large NFS clusters at that time so I can't speak what people were running. However the only other things I remember from 90s was Irix and Tru64 which was by the way my first OS ( MicroVAX 3100). Irix indeed was a very serious player in 90s but we didn't have a single SGI machine by 2003.
 
You must log in or register to reply here.
Back
Top