Slashdot Mirror
High-Performance Linux Clustering
An anonymous reader writes "High Performance Computing (HPC) has become easier, and two reasons are the adoption of open source software concepts and the introduction and refinement of clustering technology. This first of two articles discusses the types of clusters available, uses for those clusters, reasons clusters have become popular for HPC, some fundamentals of HPC, and the role of Linux in HPC."
*ducks*
With Linux and other freely available open source software components for clustering and improvements in commodity hardware, the situation now is quite different. You can build powerful clusters with a very small budget and keep adding extra nodes based on need.
Yea, I'd like to build one but I'm not sure what I'd use it for. Does that mean I'm a geek?
Bradley Holt
We spent $849,000 on an Itanium cluster and have recently found ourselves SOL since it's a dying architecture.
You can't even run Java on them.
If you "get" pointers add me as a friend (116)!
Okay, so I'd really enjoy trying something like the clustered model, just for academic kicks, but a relevant question comes to mind, at least for me.
:)
Where do people get the commodity systems cheap enough to be able to play around with this? I hardly want to spend two thousand bucks on some old P2s just to play around. Anyone have some hot tips where you can find real cheap (dare I dream... free) commodity systems to build a low-end cluster for kicks?
Also, I'm a Windows guy by trade. Will making a Linux cluster make me instantly cool?
From everything I've seen, MOSIX is having some issues right now. Unfortunately, MOSIX is one of the easiest, most flexible ways to set up an HPC, and ever since they forked, development has been slow. I did research about 2 months ago to look into setting up a small MOSIX cluster with a few computers. My main goal was to get my feet wet in setting up a cluster using a few desktop and laptop computers. I figured that setting up a cluster with my Athlon 64 x2, Athlon 64 3500+, and a few laptops would speed up compile times by quite a bit. But, it appears that the 2.6 version of MOSIX is still beta and won't support the kernel I need for my Athlon 64 x2 (versions before 2.6.9 don't support powernow with the x2, and also tend to be flaky). So, I have the choice of running a cluster with slower PC's, or waiting for better support. If you look at the year on some of those whitepapers, only one was written this year, and I'd be willing to bet they are describing how to use MOSIX with the 2.4 kernel, not 2.6. I finally gave up on the idea, as running the latest kernel is more important to me.
You can't even run Java on them.
What do you mean? I thought 1.4.2 and up had support for Itanium. Check this white paper (search for Itanium). Are their claims false, or are you running and older version of the JRE?
I'm wanting to build one of these, but I really don't need it. Time may change that.
DMCA, Hollings, Palladium. What might have sounded like paranoia is now common sense.
For some very good information on F/OSS based clustering, check out aggregate.org. They have really neat ideas, that are reasonably well doccumented and freely implementable/usable. I built a little cluster (AFAPI on a WAPERS switch) with them for my highschool senior project, and it was a great experence.
Though mpp's are kind of like clusters, and the boundary between the two is vague, I think there's definately a distinction. In many MPPs, nodes share access to memory, just at a performance penalty. Often the scientific binary is written using a message-passing tool like MPI, but the OS is often run with direct memory access. Definately from a systems-administration point of view, an mpp is different from a cluster. In an MPP you don't have 4000 root hard drives and 4000 power supplies to replace when they break. An mpp may be like a (fast) cluster from the programmer's point of view, but they are a lot simpler to deploy and manage. (Blue Gene, xt3, altix)
I also contest some of the distinctions drawn about vector processor systems. The two vector systems currently on the market, the cray X1 and the NEC SX-8 are clusters. Each node just happens to be a vector-smp. The earth simulator is a 640 node cluster of 8-way SMP boxes, where each of the processors in the smp is a vector cpu. However, the predominant programming method even on these boxes is with explicite message passing like MPI. Co-array fortran and Unified Parallel C are faster, but slow to catch on.
Good summary of the common case though.
Zero CPU implementations of this.
But their links could at least have mentioned OSCAR http://oscar.openclustergroup.org/ or my personal favorite, ROCKS http://www.rocksclusters.org/, as these are more prevalent than xCat systems.
Personally, I like Rocks, as I ran three parallel architectures (i386/AMD64/IA64), on the same based distribution, just with each tuned to their particular processor. Comes with SGE and Myrinet support out of the box, and there are Rolls, i.e. custom software assemblages, for OpenPBS, for those who prefer it, as well as PVFS. It's easy to set up, and easy to administer, as the nodes are presumed to be interchangeable and disposable. When you reboot a node, it's obliterated and a fresh OS and supplementary package repository are laid down on a clean disk. No questions about version skew.
They now have a custom roll to help you build a visualization wall, but I never had a chance to try that one. (try convincing your boss that you want 4 digital projectors and a big room to play with)
The downside to the above distributions are that they presume batch-queue environments, which is appropriate for most of my work, but less so for many people trying to simulate owning an SMP, without paying SMP prices.
Other people assure me that the current version of OSCAR is solid as well, but they seem to lag in the multiple architecture support area (Itanium is always behind), and don't current support AMD64 natively. On the other hand, they build on top of several RedHatish linuces, as opposed to Rocks where you get Centos (RHEL), period.
the more accurate the calculations became, the more the concepts tended to vanish into thin air. R. S. Mulliken
Rocks has a great system for making high-performance clusters from similar machines. A Rocks cluster consists of a front-end ("master") node and a bunch of compute nodes (and I think special-purpose nodes).
The master gets a full Linux (RedHat-based) install. It's a NFS/DHCP/Kickstart server for the compute nodes, and runs whatever other services you want the compute nodes to use. The master has two network cards and acts as a firewall (NAT optional).
The compute nodes boot via DHCP and Kickstart, downloading their kernel and whatever other OS files you want to their local disk. You decide how much NFS or local disk to use.
Job queueing is handled by, e.g., Sun Grid Engine (an Open Source queueing package) or some other queueing software.
Here's the neat thing: to make a change to a compute node setup, you change the Kickstart config and reboot all the compute nodes (as they finish whatever queued work they're doing, or immediately if you want). That makes the sysadmin's life easy, while still maintaining the speed of having the OS on the local disk.
Raise your children as if you were teaching them to raise your grandchildren, because you are.
Yeah, ok, I'm confused. I naturally thought you were talking about running java application servers or the like on Itaniums. It didn't cross my mind that you bought nearly a million dollars in hardware to run applets in your browser.
I have deployed several clusters throughout the years, mainly for research in academic environments and small companies, and I can say that clustering makes a lot of things soo much easier.
Diskless SSI clustering makes maintainance a breeze, and ensures that all systems are always in sync and up to date. All nodes can run the same system image, whether they are servers, dedicated compute nodes, or regular desktop machines.
Of course you can still have local hard disks if you want, and for some apps it is recommended, but the system boots from the servers nontheless.
OpenMosix dynamic distribution makes it possible to use heterogenous hardware, and handles highly dynamic computational load quite well. The applications just wander off to whatever physical machine will run them the fastest.
This also makes simple parallel implementations of code a lot simpler, just fork and forget, and you will pay a small overhead for the benefit of having good load-balancing automagically.
Dymanic distribution also makes it possible to use regular desktops as cluster nodes along with the dedicated compute nodes.
Need windows dualboot on some nodes? no problem, when you shut them down do boot windows, the processes that used to run on those machines just migrate to another node. When you go back to linux, processes come back.
Need explicit parallelism? no probs, MPI / PVM etc works fine together with the dynamic distribution and complements it for applications that are already well parallelized.
Scaling? This has never been an issue as long as the network infrastructure is up to speed. A decent 100mb or gigabit system has proven to be good enough for just about everything I've seen.
High availability? How about having several servers that can run hot or cold spare for each other, and which can function as compute nodes as well... Nice when a server MB catches fire (yes, I've had that, and lost as much as a few minutes of work time, (the time for someone to walk to the server room, unplug the smoking machine and restart a running (cold spare) backup server). Most of the people at the lab didn't even notice the hickup.)
Batch/job queues? no probs, use sun grid engine, write your own, or whatever. simple as cake.
I have mainly used gentoo linux for the flexibility and ease of maintainance and I can highly recommend it. It is all fairly simple to implement on gentoo. Just read up on gentoo system administration, pxelinux, tftp, openmosix, and whatever you feel you need to use it for.
The main problem right now is the lack of good openmosix support for 2.6 series of kernels. But I'm sure that some or all of this can be built with any or all of the other dynamic distribution systems out there.
If you have off-list questions please contact me at my nick at gmail.com.