30 Minute Interview

HPC is a tremendous opportunity

Kyril Faeov, General Manager, Windows HPC, Microsoft Corporation talked to Akhtar Pasha about how HPCS handles HPC workloads better in a clustered environment with one integrated management console to expedite application deployment

Kyril Faeov

Given Linux and UNIX’s prevalence in HPC, how will Windows HPC Server 2008 compete with free Linux in scenarios where you need hundreds or thousands of licenses for an installation?

Certainly Linux comes mostly free but technically it is important to address the Linux market and dissect things a bit with regard to market differentiation. If you look at academics and the government sector or labs, here you have personnel that typically employed for the purpose of doing research and development work. Though they have technical skills, yet it is harder for them to become system integrators and developers when necessary. Therefore, Linux becomes a collection of parts and not a complete solution. However, if you look at the commercial sector what dominates IT is really the costs of managing systems consistently. Moreover, what we hear from customers time and again is that they are looking for systems that utilize existing skill sets and that have a degree of consistency between different installations and strong support. This is where you see Red Hat making a significant amount of money charging for support. Moreover, if you look at HPC there are clusters of vendors that have a complete suite of solutions that goes beyond what Red Hat offers such DataSynapse (in risk and financial applications) that charge in some cases thousand of dollars per node or hundreds of dollars per CPU for traditional value on top of that. Therefore, you see there is a lot of spending on the system management stack in the HPC space.

Although HPC deployments have been on Linux, customers have always used a third party HPC workload management tool along with Red Hat. Businesses are facing the heat of skill sets that have not been available on the Linux platform and that there are not many Linux administrators. Quite a bit of money is spent on HPC workload management tools. This is because of the fact that fundamentally the entire concept operates on the premise that if you solve a big challenge they will be willing to pay for it. The challenge is not in producing a software application that powers independent nodes in the cluster, but how do you make sophisticated, advanced technology that makes a large collection of systems operate as one unit and make it high performance, reliable in a way that integrates the entire life cycle of IT and integrates with the rest of the system.

The other question is one of application development. There is incredible pressure on business folks that are supporting traders to get an application out in a matter of days certainly not more than a week that require incredible developer productivity and tools that easily allow you to test the applications and workstations and get them to the cluster at the push of a button. With the Linux environment today, you do not have that kind of productivity, which is why people are choosing the Microsoft platform. One of the biggest challenges that customers face today in adopting HPC solutions lies in the deployment of clusters and nodes. It is not hard to deploy a four-node cluster. Nevertheless, when you are deploying clusters of a few hundred nodes or a few thousand nodes, manual installation is tedious and prone to errors. Customers are looking for speedy deployment of their applications that are being developed so that they can cut the time required for deployment. At one stroke system administrators can control the servers, allocate users according to the application and workload. It automatically installs the nodes.

How does HPC Server 2008 differ from Windows Server?

There are three distinct areas of improvement in Windows HPC Server 2008 (HPCS) over its predecessors. It improves the productivity of systems administration and cluster interoperability, helps in rapid HPC application development through integration with Visual Studio 2008 and lets you seamlessly scale from workstation to cluster.

If you have a thousand node cluster, you can install either HPCS or Red Hat, on thousands of servers. There is nothing built into the Red Hat product that allows you to look at this collection of servers as one system and this is where clustering middleware comes into play. It allows administrators to take control of these collections of servers as one unit—to be able to manage it, to be able to allocate [server] resources for the users and to be able to write applications that treats it as a single supercomputer. On the management side, it allows you to define one image and then multi-case automatically and store all nodes in the cluster in parallel and within an hour you will have 200-300 nodes all provisioned in the exactly same way visible in the management console. The new administration Console based on System Center UI framework integrates every aspect of cluster management. Node grouping allows administrators to categorize and perform batch operations on compute nodes. It has built-in system diagnostics and cluster reporting that allows for job scheduling analysis using external database tools like SQL Server Analysis Services.

The next differentiator is the Job scheduler. It queues jobs and their associated tasks and allocates resources to these jobs; initiates tasks on the compute nodes of the cluster; and monitors the status of jobs, tasks, and compute nodes. The HPCS Job Scheduler has been improved to scale and support much larger cluster installations and a greater number of simultaneous jobs. It now includes new policies for greater flexibility and resource utilization, and is built to address both traditional batch jobs as well as newer service-oriented applications. It is faster, more flexible, and provides support for heterogeneous clusters, clusters whose nodes have different hardware-software configurations. The Job Scheduler helps users schedule jobs, allocate resources needed for a job, and change the tasks and properties associated with it. It includes built-in parametric support and custom job filtering, and supports heterogeneous and multi-core clusters.

The NetworkDirect, a new Remote Direct Memory Access (RDMA) network interface provides dramatic performance improvements for MPI applications running over high-speed fabrics. The shared memory Microsoft Message Passing Interface (MS-MPI) helps in the implementation of multi-core servers. Similarly, iSCSI SAN support in Windows Server 2008 and new parallel file system support and vendor partnerships help support clusters with high performance storage needs.

Since HPC clusters are popular with a broad range of mainstream users for mission-critical applications, security and integration with the existing infrastructure are essential. Windows HPC Server 2008 uses Active Directory to guide role-based security for all cluster jobs and administration. The scheduler runs each job under the context and credentials of the submitting user, not as a super user or administrator.

Is your licensing model for HPCS based on the number of cores or nodes?

If you are worried about the HPC performance and reliability, we have a very price competitive product in the form on HPCS 2008 which is quite an inexpensive product if you look at what a commercial Linux stack will look with all the pieces required and in comparison to the overall purchase price of the system or purchase price of that application. In the academic sector we always had special pricing where Windows software is extremely cost-effective or in many case completely free for example the Microsoft Software Developer Network is the collection of core Microsoft products that are frequently used by computer science departments to run their infrastructure and do to research. We have priced HPCS servers competitively for the academic sector at $50 per node. However, for the commercial user, it would be $450 per node depending upon the volume and number of users and the relationship that they have with us.

Does HPCS 2008 enjoy broad industry support?

Windows HPC Server 2008 provides a platform for independent software vendors (ISVs) and an expanded playing field for original equipment manufacturers (OEM) to enable innovation in high performance computing. AMD, Dell, HP, IBM and Intel are our hardware partners and continue to drive high-performance computing further into the mainstream. On the software front, Matlabs is our partner. Milliman (for its MG-ALFA) is our partner in the insurance sector and Ansys is another partner for Windows HPC Server 2008.

What are the killer apps for Windows HPCS?

In the public sector—academics and the government have been the traditional users of HPC. Engineering, automotive, aerospace are sectors that have used Windows workstations and using HPC Server 2008 is a natural extension for them. There are areas where it will emerge such as digital media for movie rendering, creating special effects and the like. This is an area that we are particularly excited about in India, one where we have seen early customer traction. For example, Pixar’s rendering software is a huge success on HPC.

How important is success in HPC to Microsoft’s overall business?

One of the reasons why we got into the HPC business is that we saw a tremendous opportunity for Microsoft as HPC contributes 12-15% of the server market and is one of fastest growing segment within this segment. Until we entered this market, HPC was primarily on Linux and UNIX, so it is a green-field opportunity for us to grow our server market share. Additionally HPCS 2008, along with other Microsoft applications, servers and tools, will take HPC mainstream. Our goal is to make it a part of mainstream computing, make it available to companies that could previously not afford it.