In our last post, we talked about a new type of performance race in the information technology industry. Rather than top speed or 0–60 acceleration of our computers, we’re moving to a new era where throughput, real-time accessibility and efficiency are the new pinnacles of achievement. This blog post looks at a customer example where the intelligent application of technology can make a real-world difference to a topic which has touched many of us in our every-day lives — research into cancer treatment.
One of the most common methods of cancer therapy is Photon radiation — patients are exposed to a radioactive source for a short time, using the resulting photon beam to treat cancer. The dose must be carefully calculated in order be effective, and this is one of the many workloads which are studied closely by research scientists.
There are many different ways to model this problem mathematically, but a particularly effective method is known as a Monte Carlo (MC) simulation. This type of algorithm essentially uses repeated sampling to attempt to arrive at a statistically accurate solution. You may have employed a similar computational method yourself — for example, on your journey to work in the morning; every day you can make small changes to your route — an earlier train, or a different motorway junction — and compare the resulting time you arrive at the office. You will eventually converge on a desirable solution — and you can get yourself to the office at 8.55am every morning of the week.
By processing MC dose calculation workloads, researchers can quickly and efficiently calculate the photon beam therapy needed for different cancer patients. The current challenge is to get those calculations completed as close to real-time as possible. To achieve this all we need to do is give researchers access to computers in the quickest, easiest and cheapest way. Looking at the problem as a whole, this second step is by far the easiest one — but has eluded us until recently.
Part of the problem is that modern compute clusters are reassuringly expensive — to justify buying them, you need to have a constant pipeline of work to keep them busy. Even when used as a diagnostic function, you might be able to squeeze 20 patients into a 9am-5pm consulting weekday — but then the cluster could sit idle all night and during the weekend. To compound the problem, data about patients must be carefully protected — security legislation in many countries mandates that such data is held on sovereign soil. Few hospitals have space for rooms of computers between patient beds.
In 2016, The Institute of Cancer Research (ICR) and The Royal Marsden NHS Foundation Trust in London published a paper [1] trialling the use of Amazon Web Services (AWS) public cloud to deliver near real-time MC dose calculation for a few US dollars per hour. Anonymised data was used in the trial, which was run using instances based in the AWS eu-west-1 region, based around Dublin, Ireland. The paper demonstrated that the solution, “delivers excellent linear scaling for high-resolution datasets with absolute runtimes of 1.1 to 10.9 seconds for simulating a clinical prostate and liver case up to 1% statistical uncertainty”.
Working closely with ICR, AWS and Alces have been able to demonstrate solutions to further improve the execution of this workflow. The new AWS London region (eu-west-2) will allow UK institutions such as ICR to process data in their home country, with security certification to the necessary clinical levels planned for 2017. AWS is also building further new regions around the world, broadening the options for researchers from many nations. In November 2016, AWS also announced their new Elastic GPU service, along with new instance types with embedded FPGA devices — both of these technologies open the door for researchers to develop new methods of running MC calculations more efficiently.
The challenge of delivering MC photon dose calculation workloads in near real-time is aided by using Alces Flight to auto-scale ICR’s compute clusters. This enables the research team to run their calculations ephemerally, scaling their clusters to fit the needs of the job and automatically shutting down to ensure resources are consumed cost-effectively. Through our long relationship with ICR, Alces was not only able to work directly with the HPC software and techniques, but were also in a position to integrate the workflow into the Flight Solo product available in the AWS Marketplace.
By these means Alces was able to demonstrate the ability to run MC photon dose calculation workloads in the new AWS London region, using the spot market for both CPU and GPU resources. As well as making the working method available for world-wide researchers to use in their local AWS region, launching a compute cluster with Alces Flight using spot instances can reduce operating costs by as much as 64%.
By simplifying access to and reducing the cost of running compute clusters, Alces Flight and the AWS public cloud platform are helping ICR in their goal to improve cancer detection and treatment rates, and lowering the barriers to efficient compute processing for supporting workloads. Although this just one piece of all ICR continues to achieve, the intelligent application of technology can play an active part in improving the lives of cancer patients and their families across the world.
[1] Towards real-time Photon Monte Carlo Dose Calculation in the Cloud; Peter Ziegenhein, Igor N. Kozin, Cornelis Ph. Kamerling and Uwe Oelfke; Joint Department of Physics at The Institute of Cancer Research and The Royal Marsden NHS Foundation Trust, London, UK SM2 5NG; E-mail: Peter.Ziegenhein@icr.ac.uk