Two recent announcements from Amazon and from Mellanox revealed to me that one important component in the cloud seemed to be neglected in the public discussion so far: the network. And Ed Walker’s NAS parallel benchmark came to mind which in 2008 demonstrated that message-pass¬ing latencies and bandwidth were an order of magnitude inferior between EC2 compute and compute nodes on an NCSA cluster with InfiniBand. His conclusions then were that substantial improvements can be provided to the HPC Cloud community if a high-performance network provisioning solution can be devised for this problem. In the light of his invited talk at ISC Cloud’10 in October this year in Frankfurt, I decided to ask Gilad Shainer from Mellanox a few questions, to get an update on this Cloud networking issue.
Wolfgang: Is HPC in the cloud a feasible option for real applications in industry and research?
Gilad: First we should ask if HPC in the cloud is feasible. For private clouds the answer is yes, and this is not a new concept. Many HPC systems are already being used by many users, which is what private clouds are all about. Public clouds today are not really a good match for HPC applications – high latencies, overheads, low efficiencies and utilization factors. It does not mean that future public clouds cannot meet HPC requirements, but they need to be designed with high-performance in mind.
High-performance clouds can definitely serve low performance applications as well, but vice versa does not really work. As for this question, with performance capabilities that meet the needs of the most demanding applications, HPC clusters are ideally suited for applications such as climate research, molecular modeling, physical simulations, cryptanalysis, geophysical modeling, automotive and aerospace design, financial modeling, data mining and more. Bringing the same architecture to the cloud is definitely feasible and HPC Clouds can be a very cost-effective solution for researchers who do not have the funding or IT experience to obtain and maintain their own systems. By providing on-demand high-performance cluster access to a large pool of academic and commercial researchers, HPC clouds can be a significant tool in accelerating worldwide research.
Wolfgang: How important is the networking within HPC clouds?
Gilad: How important is networking for HPC systems? Very important. There are 3 elements that need to be balanced – CPU/GPU, memory and interconnect. High-performance Cloud computing is not different and is all about efficiency and requires a simple way to deploy and access everything from single system to large scale IT resources, on demand, in real time and at an affordable cost. By utilizing networking solutions that deliver 40Gb/s throughput, 1 micro second server-to-server or storage latency, offloading and sophisticated I/O virtualization capabilities, cloud providers can perform system provisioning, workload migrations, and support multiple user requests faster and in the most efficient way, all while bringing the best ROI for the cloud providers and users.
Those items by the way are also important for non-HPC clouds. You still want to be able to host as many users as possible, so fast migrations and provisions are critical for every cloud. This is why choosing the right network is critical for the success of next generation clouds. In our case, InfiniBand connectivity makes high-performance compute and high-capacity storage available to anyone instantly, and enables HPC applications to achieve significant performance improvements, enabling researchers to get faster results.
Wolfgang: I just read the article about Mellanox’ InfiniBand solution for the University of Cambridge. How important was to use InfiniBand there?
Gilad: Incorporating Mellanox 40Gb/s InfiniBand, with its GPU efficiency and offloading capabilities, as the clustering backbone for the University of Cambridge’s Solutions Centre was essential to their cloud provision requirements. Their new HPC Cloud cluster provides on-demand access to support over 400 internal users spread across 70 research groups ranging from traditional sciences such as chemistry, physics and biology, to rapidly growing areas for HPC-based research such as bio-medicine, clinical-medicine and social sciences.
Wolfgang: What is Mellanox’ cloud computing strategy in general?
Gilad: Mellanox server and storage interconnect solutions have been designed to deliver the needed networking capabilities of bandwidth, latency and virtualization offloads for the Cloud infrastructure. The company’s end-to-end InfiniBand connectivity products are quickly becoming the networking products of choice for cloud computing infrastructures. The growth is due to the inherent benefits InfiniBand provides in optimizing performance for technologies commonly used in the Cloud, e.g. server virtualization, multi-core CPU servers, and high-performance data storage. These benefits span industries and applications, private and public clouds.
Wolfgang: How do you see the role of clouds in the future of HPC?
Gilad: With performance capabilities that meet the needs of the most demanding applications, Clouds can be ideally suited for a broad array of research services and market applications. Furthermore, if you look on the next generation HPC systems, those systems will create an exaflood of data, and you will need the capability to host the data, analyze the data and be able to have the data shared with many researchers and scientists around the world. HPC clouds will be the most suitable solution for that. By the way, I am also involved with the HPC Advisory Council, and there we also explore usage models of HPC in cloud computing and future direction. You can see the publications we made at http://www.hpcadvisorycouncil.com/subgroups_hpc_cloud.php.
Wolfgang: And what is the main theme of your talk at ISC Cloud’10 end of October in Frankfurt?
Gilad: Clouds can be seen as a superset of data centers and HPC systems, or maybe as a new environment with its own requirements. The main theme of my presentation will revolve around the latest developments in the networking world (both Ethernet and InfiniBand) and how these developments are helping make clouds more efficient and accessible, and I will highlight various examples of usage cases and simulation results.
Dr. Wolfgang Gentzsch is the General Chair for ISC Cloud’10, taking place October 28-29, in Frankfurt, Germany. ISC Cloud’10 will focus on practical solutions by bridging the gap between research and industry in cloud computing. Information about the event can be found at the ISC Cloud event website. HPC in the Cloud is a proud media partner of ISC Cloud’10.