Introduction to High Performance Computing (HPC)
Last edit: dombennett (07 Oct 2016 18:32) | Revisions: 7 | Created by: dombennett | Rating: 0
Introduction
The HPC (High Performance Computing) service is a means to run programs in parallel. Although this can be done on your local machine, it cannot be done on anywhere near the same scale. This allows you to run programs that might have taken months if not years to complete in a few hours or days. The HPC is a UNIX based server that consists of multiple CPUs. An HPC service is usually provided by your university or research institute. If not, there are other commercial services available. These servers are often composed of 1,000s of cores each equivalent to a powerful computer.
What is parallel computing?
Parallel computing is the opposite of sequential computing. Instead of running a program over a set of data sequentially, you split the data up into chunks and run the program sequentially over those chunks simultaneously. To give a trivial example, imagine a program that consists of a single for loop that prints the numbers 1 to 100. A simple way to parallelise this program would be to run two for loops simultaneously, one counting the numbers 1-50 and the other, 51-100. This, in theory, should be twice as fast. The only downside is running two processes in parallel will take twice the amount of computing power.
Do I need the HPC?
If you have a program that is taken a long time you may need to use the HPC. I would suggest, however, to only use the HPC if you've got a program which is taking months to complete rather than days. Understanding how to use the HPC can be a long process and for programs that can be parallelised on your local machine there are simpler alternatives. If you're an R user here is a really good blog post on how to parallelise your for loops. For python users, you can make use of a variety of libraries: Queue, threading and subprocess. And then in this example.
A simple guide on how to use the HPC
You will need
- A user account on an HPC server. This will usually require a demonstration of need, competence in using UNIX command-line (see our guide) and, often, for you to complete a tutorial on using the HPC (institution dependent).
- A program that can be parallelised.
How to parallelise a program
In the above trivial example, we had a for loop counting from 1-100. This is a readily parallelisable program because the process of printing each number is independent. We can cut the data set (in this case the numbers 1-100) into a maximum of 100 chunks, and run each chunk as a separate process. The process of printing a number is called a job. So long as jobs are independent then a program can be easily parallelised. If, however, our program was more complicated and the results of one job depended on the results of another then the program would be very difficult to parallelise. Parallelising such programs is possible but is beyond the scope of this document (look up the map-reduce method and perhaps find an expert in high performance computing).
Make sure the program you are using has options or a version that can run on an HPC server, e.g. look for the words 'MPI'. Otherwise, if you're running your own program then you may have to rejig it to get it to work on the HPC. You will need to check the following things:
- Is my program divisable into independent jobs?
- How can I cut my program up into these jobs?
- How can I run each job independently?
If you can run a script with an argument that will work on one part of your data set independently of other parts, then you can run your program in parallel. At this stage you simply need to create a task farm script that will run your program in chunks. Below is a python script that outlines how this can be done. The script consists of three functions: `job()`, `master()` and `worker()`. It takes a worklist (the list of all jobs that needs to be run), feeds them to a worker (as executed by the `worker()` function) that executes the job, waits for the job to finish and sends a signal to the master (as executed by the `master()` function) to call for another job. This continues until the worklist is depleted. The job can be anything (excution of python scripts, R scripts or any other program), it will simply run a script using whatever argument has been given to it via the worklist.
#! /bin/usr/env python # PACKAGES import os import sys from mpi4py import MPI # GLOBALS comm = MPI.COMM_WORLD rank = comm.Get_rank() # FUNCTIONS def job(jobid): '''Run independent job''' print "running job ", jobid def master(worklist): '''The task master takes a work list and farms it out to the workers.''' # Check we are not using more ranks than what we need. nranks = comm.Get_size() if len(worklist) < (nranks-1): print "WARNING: There are fewer tasks than ranks. You are wasting \ resources!" for rank in range(len(worklist)+1, nranks): # Catch the send buff = comm.recv(source=rank) # Send completion signal comm.send("", dest=rank) # Reset number of ranks. nranks = len(worklist)+1 # Farm out tasks for task in worklist: # Probe for any idle worker. status = MPI.Status() buff = comm.recv(source=MPI.ANY_SOURCE, status=status) idle_worker = status.Get_source() # Give worker task comm.send(task, dest=idle_worker) # Send null task to signal completion. for rank in range(1, nranks): comm.send("", dest=rank) return def worker(task_operation): '''The worker takes as an input the operation it should apply to the tasks that it receives from the task master.''' while True: # Alert master that we are waiting. comm.send("") # Wait for a task task = comm.recv() if task == "": break else: task_operation(task) return if __name__ == '__main__': worklist = sys.argv[1:] # obtain worklist from command-line print 'Using: ', worklist if rank == 0: master(worklist) else: worker(job)
For example, to go back to my trivial example of counting 1-100. Let's imagine I want to run two workers (1-50, 51-100). I could then create two R scripts that looped through the two sets of numbers: 1_50.R and 51_100.R. The job() function could then be a function that takes as an argument the name of an R script and executes it. I would then execute the task farm like so: task_farm.py 1-50.R 51-100.R. (In practice, this is not a great example as it would require each job to have its own script.)
PBS script
The PBS script is the instructions for running your parallel program across all the nodes and clusters available to you through the HPC server. It specifies the number of processes you wish to use and the amount of time you require for it to run. Below is an example of a pbs script.
#!/bin/bash
#PBS -l select=1:ncpus=16:mpiprocs=16:ompthreads=1
#PBS -l walltime=24:00:00
mpiexec python path_to_task_farm.py argument_for_task_farm.py
In this example it consists of two key elements: the PBS instructions denoted with `#PBS` and the execution command for the task farm script. The PBS instructions select 1 node of 16 CPUs. It allows up to 16 workers (called mpiprocs) and each of these workers use 1 thread each. If I wanted to double the number of workers I would make select=2, as this would select 2 nodes instead of 1. The instructions also prevent the program from running for more than 24 hours. You should check with your administrator what sort of cores are available to you and what the maximum walltime is.
HPC structure
An HPC server contains multiple cores. Each core can have a number of CPUs. For example, on my HPC server I have access to three types of core: 16CPU, 12CPU and an 8CPU. So if I wanted to select 24 CPUs I could select 2 12CPU cores, or if I needed 32 I could select 4 8CPU cores or 2 16CPU cores. Just make sure you are not selecting cores that do not exist on the server. If you do this, your program will be queued indefinitely. (On my system it is put on a queue called Monster.)
Launching and checking
To launch a PBS script you use the command qsub:
qsub my_pbs_script.sh
This will add the script to a queue. To check whether your script is being run
use the `qstat` function.
qstat # no arguments necessary
This will print to screen, the ID of your job, how long the program has been running or whether it is still in the queue. All output and error messages generated by your program are sent to files called: [name of job].e[ID] and [name of job].o[ID]. These will not be available until after walltime, and there is always a few minutes delay. It is always good therefore to have an alternative way of making sure your program is running correctly. For programs jobs always make changes to directories, I can then check the progress of my program by counting the number of changes to these directories.
HPC directory house-keeping
Logging in
To log in you will need to use the ssh command or WinSCP if you're using Windows. You will need to check what your account username and password is through the system administrator. Once logged in you will be taken to your home directory. A HPC server consists of multiple nodes, you log into the login node. Usually you will not be running any programs on this login node. It is just the portal for executing commands that run across all the nodes in the HPC.
Home and work directory
You usually have available to you two directories where you can store data: the home and work directories. The home directory is smaller than the work, you should check these details with your administrator. I usually store my programs in the home directory and all my data and results files in the work directory. I also create a link to my work directory in the home directory to prevent having to write the work directory multiple times, e.g.
ln -s /work/[username] work
Modules
You usually have a variety of programs available to you already on the HPC server. If a command is not recognised, you may have to manually load programs using the module function.
module av # list all available modules
module load [name-of-module] # load the module
I add these commands to my bash profile for any programs I use regularly.
# make sure you're in your home directory
echo 'module load python/2.7.3 >> .bash_profile' # my preferred version of python
Installing packages and programs
You may have to install dependent packages, libraries and programs in order to run your program(s) that may not be available from modules. You can either request an administrator to install it for you, or you may have to compile it yourself in your home directory. You can compile programs yourself so long as you do not append them to the system path (which requires administrator privileges). Instead you will have to directly call your compiled version of the program with its absolute link. Or you could create your own virtual environment (document coming for this soon!). When compiling yourself make sure you are aware of the HPC architecture (Intel, AMD …) and call the relevant modules before compilation, this will allow the compiled program to run more efficiently.
Useful links
Your best source of information for using the HPC server is from your administrator and the HPC technical team. They should offer courses and guides. Otherwise, there is plenty of information on the web that can be of use. Here are some links for different institute's HPC services (some institutes are better at giving instructions than others ….). Although HPC systems differ between institutions they are quite similar and can still be helpful, particular if common problems keep occurring.
And background information ….
Contributors
Dom Bennett (plus thanks to Gerard Gorman for the python task farm example)
I think this article confuses cores and nodes. A node is a vm in the cluster that contains multiple cpus. Cores are analogous to cpus