([email protected])
This document describes how we set up our Linux computing clusters for high-performance computing which we need for our research.
Use the information below at your own risk. I disclaim all responsibility for anything you may do after reading this HOWTO. The latest version of this HOWTO will always be available at http://www.ram.org/computing/linux/linux_cluster.html.
Unlike other documentation that talks about setting up clusters in a general way, this is a specific description of how our lab is setup and includes not only details the compute aspects, but also the desktop, laptop, and public server aspects. This is done mainly for local use, but I put it up on the web since I received several e-mail messages based on my newsgroup query requesting the same information. Even today, as I plan another 64-node cluster, I find that there is a dearth of information about exactly how to assemble components to form a node that works reliably under Linux that includes information not only about the compute nodes, but about hardware that needs to work well with the nodes for productive research to happen. The main use of this HOWTO as it stands is that it's a report on what kind of hardware works well with Linux and what kind of hardware doesn't.
This section covers the hardware choices I've made. Unless noted in the known hardware issues section, assume that everything works really well.
Hardware installation is also fairly straight-forward unless otherwise noted, with most of the details covered by the manuals. For each section, the hardware is listed in the order of purchase (most recent is listed first).
32 machines have the following setup each:
32 machines have the following setup each:
32 machines have the following setup each:
32 machines have the following setup each:
32 machines have the following setup each:
Two servers for external use (dissemination of information) with the following setups:
We have identified at least two kinds of users of our cluster: those that need (i.e., take advantage of) permanent local processing power and disk space in conjunction with the cluster to speed up processing, and those that just need only the cluster processing power. The former are assigned "desktops" which are essentially high-performance machines, and the latter are assigned dumb "terminals". Our desktops are usually dual or quad processor machines with the current high-end CPU being a 1.6 GHz Opteron, having as much as 10 GB of RAM, and over 1 TB of local disk space. Our terminals are essentially machines where a user can log in and then run jobs on our farm. In this setup, people may also use laptops as dumb terminals.
We generally use/prefer Viewsonic monitors, Microsoft Intellimouse mice, and Microsoft Natural keyboards. These generally have worked quite reliably for us.
For visual access to the nodes, we initially used to use KVM switches with a cheap monitor to connect up and "look" at all the machines. While this was a nice solution, it did not scale. We currently wheel a small monitor around and hook up cables as needed. What we need is a small hand held monitor that can plug into the back of the PC (operated with a stylus, like the Palm).
For networking, we generally use Netgear and Cisco switches.
Our vendor is Hard Drives Northwest ( http://www.hdnw.com). For each compute node in our cluster (containing two processors), we paid about $1500-$2000, including taxes. Generally, our goal is to keep the cost of each processor to below $1000 (including housing it).
The following kernels and distributions are what are being used:
These distributions work very well for us since updates are sent to us on CD and there's no reliance on an external network connection for updates. They also seem "cleaner" than the regular Red Hat distributions, and the setup is extremely stable.
We use Shorewall 1.3.14a (( http://www.shorewall.net) for the firewall.
We use our own software for parallelising applications but have experimented with PVM and MPI. In my view, the overhead for these pre-packaged programs is too high. I recommend writing application-specific code for the tasks you perform (that's one person's view).
Linux and most software that run on Linux are freely copiable.
This section describes disk partitioning strategies. Our goal is to keep the virtual structures of the machines organised such that they are all logical. We're finding that the physical mappings to the logical structures are not sustainable as hardware and software (operating system) change. Currently, our strategy is as follows:
farm/cluster machines:
partition 1 on system disk - swap (2 * RAM)
partition 2 on system disk - / (remaining disk space)
partition 1 on additional disk - /maxa (total disk)
servers:
partition 1 on system disk - swap (2 * RAM)
partition 2 on system disk - / (4-8 GB)
partition 3 on system disk - /home (remaining disk space)
partition 1 on additional disk 1 - /maxa (total disk)
partition 1 on additional disk 2 - /maxb (total disk)
partition 1 on additional disk 3 - /maxc (total disk)
partition 1 on additional disk 4 - /maxd (total disk)
partition 1 on additional disk 5 - /maxe (total disk)
partition 1 on additional disk 6 - /maxf (total disk)
partition 1 on additional disk(s) - /maxg (total disk space)
desktops:
partition 1 on system disk - swap (2 * RAM)
partition 2 on system disk - / (4-8 GB)
partition 3 on system disk - /spare (remaining disk space)
partition 1 on additional disk 1 - /maxa (total disk)
partition 1 on additional disk(s) - /maxb (total disk space)
Note that in the case of servers and desktops, maxg and maxb can be a single disk or a conglomeration of disks.
Install a minimal set of packages for the farm. Users are allowed to configure desktops as they wish, provided the virtual structure is kept the same described above is kept the same.
I believe in having a completely distributed system. This means each machine contains a copy of the operating system. Installing the OS on each machine manually is cumbersome. To optimise this process, what I do is first set up and install one machine exactly the way I want to. I then create a tar and gzipped file of the entire system and place it on a bootable CD-ROM which I then clone on each machine in my cluster.
The commands I use to create the tar file are as follows:
tar -czvlps --same-owner --atime-preserve -f /maxa/slash.tgz /
I use a script called go that takes a machine
number as its argument and untars the slash.tgz file on the
CD-ROM and replaces the hostname and IP address in the appropriate
locations. A version of the go script and the input files for
it can be accessed at:
http://www.ram.org/computing/linux/linux/cluster/. This script
will have to be edited based on your cluster design.
To make this work, I use Martin Purschke's Custom Rescue Disk
(
http://www.phenix.bnl.gov/~purschke/RescueCD/) to create a
bootable CD image containing the .tgz file representing the cloned
system, as well as the go script and other associated files.
This is burned onto a CD-ROM.
There are several documents that describe how to create your own custom bootable CD, including the Linux Bootdisk HOWTO ( http://www.linuxdoc.org/HOWTO/Bootdisk-HOWTO/), which also contains links to other pre-made boot/root disks.
Thus you have a system where all you have to do is insert a CDROM, turn on the machine, have a cup of coffee (or a can of coke) and come back to see a full clone. You then repeat this process for as many machines as you have. This procedure has worked extremely well for me and if you have someone else actually doing the work (of inserting and removing CD-ROMs) then it's ideal. In my system, I specify the IP address by specifying the number of the machine, but this could be completely automated through the use of DHCP.
Rob Fantini ( [email protected]) has contributed modifications of the scripts above that he used for cloning a Mandrake 8.2 system accessible at http://www.ram.org/computing/linux/cluster/fantini_contribution.tgz.
FAI ( http://www.informatik.uni-koeln.de/fai/) is an automated system to install a Debian GNU/Linux operating system on a PC cluster. You can take one or more virgin PCs, turn on the power and after a few minutes Linux is installed, configured and running on the whole cluster, without any interaction necessary.
SystemImager ( http://systemimager.org) is software that automates Linux installs, software distribution, and production deployment.
If you have DHCP set up, then you don't need to reset the IP
address and that part of it can be removed from the go
script.
DHCP has the advantage that you don't muck around with IP addresses at all provided the DHCP server is configured appropriately. It has the disadvantage that it relies on a centralised server (and like I said, I tend to distribute things as much as possible). Also, linking hardware ethernet addresses to IP addresses can make it inconvenient if you wish to replace machines or change hostnames routinely.
The hardware in general has worked really well for us. Specific issues are listed below:
The AMD dual 1.2 GHz machines run really hot. Two of them in a room increase the temperature significantly. Thus while they might be okay as desktops, the cooling and power consumption when using them as part of a large cluster is a consideration. The AMD Palmino configuration described previously seems to work really well, but I definitely recommend getting two fans in the case--this solved all our instability problems.
Some tar executables apparently don't create a tar file the nice way they're supposed to (especially in terms of referencing and de-referencing symbolic links). The solution to this I've found is to use a tar executable that does, like the one from RedHat 7.0.
This section is still being developed as the usage on my cluster evolves, but so far we tend to write our own sets of message passing routines to communicate between processes on different machines.
Many applications, particularly in the computational genomics areas, are massively and trivially parallelisable, meaning that perfect distribution can be achieved by spreading tasks equally across the machines (for example, when analysing a whole genome using a technique that operates on a single gene/protein, each processor can work on one gene/protein at a time independent of all the other processors).
So far we have not found the need to use a professional queueing system, but obviously that is highly dependent on the type of applications you wish to run.
For the single most important program we run (our ab initio protein folding simulation program), using the Pentium 3 1 GHz processor machine as a frame of reference, on average:
Xeon 1.7 GHz processor is about 22% slower
Athlon 1.2 GHz processor is about 36% faster
Athlon 1.5 GHz processor is about 50% faster
Athlon 1.7 GHz processor is about 63% faster
Xeon 2.4 GHz processor is about 45% faster
Xeon 2.7 GHz processor is about 80% faster
Opteron 1.4 GHz processor is about 70% faster
Opteron 1.6 GHz processor is about 88% faster
Yes, the Athlon 1.5 GHz is faster than the Xeon 1.7 GHz since the Xeon executes only six instructions per clock (IPC) whereas the Athlon executes nine IPC (you do the math!). This is however an highly non-rigourous comparison since the executables were each compiled on the machines (so the quality of the math libraries for example will have an impact) and the supporting hardware is different.
These machines are incredibly stable both in terms of hardware and software once they have been debugged (usually some in a new batch of machines have hardware problems), running constantly under very heavy loads. One common example is given below. Reboots have generally occurred when a circuit breaker is tripped.
2:29pm up 495 days, 1:04, 2 users, load average: 4.85, 7.15, 7.72
The following people have been helpful in getting this HOWTO done:
The following documents may prove useful to you---they are links to sources that make use of high-performance computing clusters: