The Evolution of “DropAndCompute” – easing the pain of accessing High Throughput Computing (and similar) facilities
In my previous blog post I mentioned that I am a member of a team that supports High Throughput Computing (HTC) at The University of Manchester via a 1600+ core ‘condor pool’. In order to make it as easy as possible for our researchers to make use of this resource one of my colleagues, Ian Cottam, created a system called DropAndCompute. In this guest blog post, Ian describes DropAndCompute and how it evolved into the system we use at Manchester today.
The Evolution of “DropAndCompute” by Ian Cottam
DropAndCompute, as used at The University of Manchester’s Faculty of Engineering and Physical Sciences, is an approach to using network (or grid or cloud based) computational resources without having to know the operating system of the resource’s gateway or any command line tools of either the resource itself –Condor in our case — or in general. Most such gateways run a flavour of Unix, often Linux. Many of our users are either unfamiliar with Linux or just prefer a drag-and-drop interface, as I do myself despite using various flavours of Unix since Version 6 in the late 70s.
Why did I invent it? On its original web site description page wiki.myexperiment.org/index.php/DropAndCompute the following reasons are given:
- A simple and uniform drag-and-drop graphical user interface, potentially, to many resource pools.
- No use of terminal windows or command lines.
- No need to login to remote hosts or install complicated grid-enabling software locally.
- No need for the user to have an account on the remote resources (instead they are accounted by having a shared folder allocated). Of course, nothing stops the users from having accounts should that be preferred.
- No need for complicated Virtual Private Networks, IP Tunnelling, connection brokers, or similar, in order to access grid resources on private subnets (provided at least one node is on the public Internet, which is the norm).
- Pop-ups notify users of important events (basically, log and output files being created when a job has been accepted, and when the generated result files arrive).
- Somewhat increased security as the user only has (indirect) access to a small subset of the computational resource’s commands.
The first version was used on a Condor Pool within our interdisciplinary biocentre (MIB). A video of it in use is shown below
Please do take the time to look at this video as it shows clearly how, for example, Condor can be used via this type of interface.
This version was notable for using the commercial service: Dropbox and, in fact, my being a Dropbox user inspired the approach and its name. Dropbox is trivial to install on any of the main platforms, on any number of computers owned by a user, and has a free version giving 2GB of synchronised and shared storage. In theory, only the computational resource supplier need pay for a 100GB account with Dropbox, have a local Condor submitting account, and share folders out with users of the free Dropbox-based service.
David De Roure, then at the University of Southampton and now Oxford, reviewed this approach here at blog.openwetware.org/deroure/?p=97, and offers his view as to why it is important in helping scientists start on the ‘ramp’ to using what can be daunting, if powerful, computational facilities.
Quickly the approach migrated to our full, faculty-wide Condor Pool and the first modification was made. Now we used separate accounts for each user of the service on our submitting nodes; Dropbox still made this sharing scheme trivial to set up and manage, whilst giving us much better usage accounting information. The first minor problem came when some users needed more –much more in fact– than 2GB of space. This was solved by them purchasing their own 50GB or 100GB accounts from Dropbox.
Problems and objections
However, two more serious problems impacted our Dropbox based approach. First, the large volume of network traffic across the world to Dropbox’s USA based servers and then back down to local machines here in Manchester resulted in severe bottlenecks once our Condor Pool had reached the dizzy heights of over a thousand processor cores. We could have ameliorated this by extra resources, such as multiple submit nodes, but the second problem proved to be more of a showstopper.
Since the introduction of DropAndCompute several people –at Manchester and beyond– have been concerned about research data passing through commercial, USA-based servers. In fact, the UK’s National Grid Service (NGS) who have implemented their own flavour of DropAndCompute did not use Dropbox for this very reason. The US Patriot Act means that US companies must surrender any data they hold if officially requested to do so by Federal Government agencies. Now one approach to this is to do user-level encryption of the data before it enters the user’s dropbox. I have demonstrated this approach, but it complicates the model and it is not so straightforward to use exactly the same method on all of the popular platforms (Windows, Mac, Linux).
To tackle the above issues we implemented a ‘local version’ of DropAndCompute that is not Dropbox based. It is similar to the NGS approach, but, in my opinion, much simpler to setup. The user merely has to mount a folder on the submit node on their local computer(s), and then use the same drag-and-drop approach to get the job initiated, debugged and run (or even killed, when necessary). This solves the above issues, but could be regarded as inferior to the Dropbox based approach in five ways:
1. The convenience and transparency of ‘offline’ use. That is, Dropbox jobs can be prepared on, say, a laptop with or without net access, and when the laptop next connects the job submissions just happens. Ditto for the results coming back.
2. When online and submitting or waiting for results with the local version, the folder windows do not update to give the user an indication of progress.
3. Users must remember to use an email notification that a job has finished, or poll to check its status.
4. The initial setup is a little harder for the local version compared with using Dropbox.
5. The computation’s result files are not copied back automatically.
So far, only item 5 has been remarked on by some of our users, and it, and the others, could be improved with some programming effort.
A movie of this version is shown below; it doesn’t have any commentary, but essentially follows the same steps as the Dropbox based video. You will see the network folder’s window having to be refreshed manually –this is necessary on a Mac (but could be scripted); other platforms may be better– and results having to be dragged back from the mounted folder.
I welcome comments on any aspect of this –still evolving– approach to easing the entry ‘cost’ to using distributed computing resources.
Our Condor Pool is supported by three colleagues besides myself: Mark Whidby, Mike Croucher and Chris Paul. Mark, inter alia, maintains the current version of DropAndCompute that can operate locally or via Dropbox. Thanks also to Mike for letting me be a guest on Walking Randomly.