01 November 2011
Expect advances in Big Data and HPC Clouds to be in the spotlight this year at SC11
Each year at the annual supercomputing show (SC11) there is always enough marketing buzz swirling around to make one dizzy. In recent years multi-core, GP-GPU, and Top500 surprises have held the market’s attention. This year, there are two topics that deserve your attention at the show. These are "big data" and HPC Cloud. There may be more, but these two are worth watching.
Since Cloud news has been raining down on us for a while (pun intended), let's take a look at the big data buzz first. The term "big data" refers to the ability to use large amounts of data efficiently. Of course, large storage arrays can hold petabytes of data, but the access times can cripple performance when compared to the speeds of modern processors and memory.
Standard databases are designed to work with slow storage and large data sets, but many scientific applications fall outside of a standard database structure. One solution to address slow disk speeds is using Solid State Disks (SSD) on the nodes to provide much faster access times in comparison to traditional spinning media. Another approach is to use Random Access Memory (RAM) -based disks (a devices that looks like a hard disks but is are actually all completely composed of RAM) from companies such as Kove. Both of these solutions help increase disk speeds, but still do not solve the need for a large memory space.
Large memory space requirements are the sorts of problems that exist in the realm that "big data" solutions address. The first "big data" examples are the traditional predictive simulations. These include areas such as structural engineering, computational chemistry, and bioinformatics. Many of these applications require as much contiguous memory as possible for large models, but don't work well in a distributed (cluster) environment. The second type of problem to be considered is large-scale data mining. For example, searching the results of several large predictive simulation models for subtle changes or results that are not perceptible to humans (i.e., finding a needle in a field of haystacks). To address these problems, there are products that combine the memory footprint of multiple servers into one large global memory.
Two of these products include ScaleMP™ (virtual SMP software approach) and Numascale (hardware interconnect approach). With either of these methods users can maintain the price-to-performance level of commodity servers while opening the door to much larger problem sizes. Both these solutions address problems traditionally reserved for very expensive SMP machines with large memory footprints. Expect to see more "big data" clusters at SC this year.
In addition, there are problems that combine predictive analysis with existing databases or combinations of existing databases. These problems represent a new frontier in HPC. Consider using a patient's symptoms and lab results as a starting point for an intelligent search of all known medical literature, which is impossible for any single human to read in its entirety, and then using the patient’s genome to design or recommend a treatment that addresses the specific ailment and minimizes side effects (under supervision of medical professionals, of course). A good example of this type of application is the Jeopardy-playing computer Watson - who will be attending SC11.
The second area to watch is cloud computing. Of course "The Cloud" has been over-hyped recently, but there are good usage cases for a cloud designed to work for HPC applications. These types of clouds are different than the traditional cloud services that are sold by many vendors, including HPC Tools. In particular, an HPC cloud needs to keep the applications closer to the hardware than most other cloud solutions.
There is also some value in understanding cloud designs and usage modes. A simple definition of a cloud is a flexible and elastic computing resource that is purchased as a service (in accounting terms it is an "expense"). Contrast this with a cluster that has a fixed number of computing elements and is purchased as a physical resource (in accounting terms, a "capital asset"). This leads one to considering the potential to treat the fixed computing element of a cluster as a private cloud provisioned resource.
Clouds can be public or private or a combination of both. Private clouds make sense for organizations that want to pool the computational budgets from various departments and create a large central private shared resource - much like mainframes of the past. A cloud offers a flexible solution that is shared (and billed appropriately) throughout the organization. Note that a Distributed Resource Manager (i.e., Grid Engine, Torque/Maui) can almost accomplish the same thing, but the underlying operating system is usually not under user control (i.e., it is flexible only to a point). Some resource mangers do have this capability and thus may even be considered "clouds." Additionally, some organizations require security, which limits the use of public clouds.
Leasing an asset is also an option, but it is still not elastic. In general, a "pay as you go" expense is often more desirable than a "pay it all upfront" expense for many organizations that have time-limited projects (i.e., more project engineering and less open-ended R&D).
A usage model that may play a larger role this year at SC11 is "HPC cloud bursting." In this approach, a local dedicated cluster may be supplemented with either a public or a private cloud. From a user standpoint, jobs will still be run through a local resource manager which will handle all of the details required for it to be able to run in the cloud or on a local resource. There are HPC cloud performance issues to consider as well, but that is another topic. HPC clouds may ultimately expand the market because they offer lower upfront costs and flexibility.
There will be plenty of other SC11 news and events worth noting this year, but both big data and HPC clouds are two areas that are worth watching and may impact the direction of HPC in the future.
