In the recent lists, the TOP500 coverage has shifted from pure high-performance computing (HPC) to include more enterprise-based solutions. In the latest TOP500 release, the 30th list, published in November 2007, the majority of the systems were enterprise datacenter (EDC)-based, mainly interconnected rack-mounted systems, while the minority of the systems were the traditional high-performance computers, mainly cluster-based solutions. In this article we will analyze the historical and current technological trends in high-performance computing and provide an updated analysis model for the TOP500 project.
“The only constant is change,” said Heraclitus, an ancient Greek philosopher. This is certainly true when describing the market of high-performance computing. The HPC market is characterized by a rapid change of architectures, technologies and usage. The only continuous, steady characteristic of HPC is the ever-growing demand for performance, showing an increase of 100X every ten years based on the TOP500 list.
The enterprise datacenter market is less tolerant to rapid changes, and typically changes are much more moderate. While the HPC market drives the technology further and evaluates many leading-edge architectures, only the proven solutions that have been widely adopted by HPC, and in particular by the commercial HPC markets, are accepted and spread into the EDC market.
The Era of Clustered Commodity Servers
Clustered commodity servers have become the dominant solution both for HPC and EDC, as they offer tremendous price/performance benefits, unparalleled flexibility in deployment, and reduced long-term maintenance. According to the latest TOP500 list, more than 80 percent of the listed systems are clusters. With the fast adoption of the cluster architecture, the importance of interconnect solutions has risen as well.
The use of off-the-shelf commodity and standard components has made its mark on the interconnect solutions, making Gigabit Ethernet and InfiniBand the dominant solutions. In the early days, when single-core CPUs were the common solution, Gigabit Ethernet was used mainly for the cases where there was no need for intensive I/O, and InfiniBand for high-performance computing or I/O intensive enterprise applications. Since the mid 2000s, multi-core processors have taken over the single-core CPUs due to the ever-increasing need for performance and increasing demand for low power solutions. This new trend pushed the need for fatter standard interconnects that can handle the increasing I/O demands, as more CPUs share the same connection.
Maximizing productivity and reducing power consumption have become the key issues in today’s compute solution. In the HPC segment, most applications utilize the entire compute resource, and therefore require high throughput and low-latency connectivity between the cluster server nodes. In the EDC segment, many applications are not compute intensive, and therefore virtualization becomes essential for increasing system productivity. Virtualization enables running multiple environments or multiple applications on the same compute system in order to maximize the CPU utilization. This solution creates the same load on the interconnect, as more throughput is required between servers and between servers and storage. While HPC and virtualized EDC environments are different from the application perspective, they require the same characteristics from the cluster interconnect.
The TOP500 List
The TOP500 project provides a list of sites operating the 500 most powerful computer systems. It does not mean that those systems are being used daily as a single supercomputer, and in some cases the daily usage is rather from a single server or a small group of servers. In those cases, the sites or the systems manufacturers gather together the site’s compute resources to form a single supercomputer to execute the LINPACK benchmark and submit the results to the TOP500 project. This action is sometimes done during installations of new systems. Since the compute resources in most cases are already connected together, the task of measuring them as a whole is an easy one.
The TOP500 list is considered an HPC-related list, and many analyze the list statistics for understanding market and technology trends. When the system architecture converged for HPC and EDC, the TOP500 list shifted away from providing statistics solely for the HPC market. In particular, this is the case with the latest lists, where clusters have become the dominant solution across the different markets. Clustering enables single servers to scale up and form a supercomputer, even if it is just for one day.
Moreover, clustering and the use of off-the-shelf components brings the power of supercomputing to the HPC masses, and to many other markets. The recent TOP500 lists (in particular the latest one — the 30th) shifted from representing only the HPC market to representing the entire cluster market, or in general, interconnected servers. In order to provide better and accurate analysis of the TOP500 list, one needs to break the list into two separate lists — the upper part, the top 100 systems, which continue to truly represent the HPC market, and the lower part, which represents the cluster market — both HPC and EDC.
TOP100: High Performance Computing
The pure HPC portion of today’s TOP500 list is the TOP100. This part is ruled by the supercomputers that actually serve as supercomputers. The TOP100 systems can be divided into two main categories: clusters with 51 percent of the entries (up from the 41 percent on the November 2005 list) and MPPs with 47 percent (down from 55 percent on the November 2005 list). As observed in the recent lists, clusters continue to show a strong growth and have become the preferred solutions for HPC systems.
The types of interconnects used in recent years clearly indicate the domination of standard interconnects over propriety solutions, which is consistent with the vast adoptions of standard and off-the-shelf components. InfiniBand has become the natural choice for HPC and in particular for clusters, connecting more than 58 percent of the TOP100 clusters. The number of InfiniBand connected clusters has shown a positive growth in every TOP100 list since its first appearance on the TOP500 list. This trend is anticipated to continue throughout the following years, especially with the increasing number of CPU cores per server node platform, which requires the highest bandwidth and lowest latency.
The use of proprietary interconnect solutions for clusters has been reduced from 70 percent in November 2005 to less than 20 percent in November 2007, indicating the market preference for standard components (which nowadays also provide superior performance and price/performance). The use of InfiniBand is not solely for clusters, since many MPP-based systems use InfiniBand interconnects as well. Another interesting point is the non-existence of 10GigE-based solutions. The main reasons for that is the superior performance and the advanced HPC-related features of InfiniBand over 10GigE, such as congestion control, adaptive routing and extremely low CPU overhead.
TOP101-500: General Purpose Clusters
While the TOP100 is divided between clusters and MPPs, the 400 systems that make up the lower segment of the TOP500 list (the systems ranked 101-400) are mostly clusters. Of these, 355 systems, almost 90 percent, are labeled clusters, while only 44 systems are marked as MPPs.
The lower segment of the TOP500 has become the cluster’s segment, and represents the general cluster market, both HPC and EDC. For HPC, this segment represents the lower end of the HPC systems, as the higher end is being represented in the TOP100. Many of the systems are not being used for single applications and some of them are “supercomputers for a day,” when the Linpack benchmark is run. The demand for standard components exists throughout the TOP500 list, and in the lower segment, InfiniBand and GigE connect almost 90 percent of the systems (347 systems out of 400) and 95 percent of these systems are marked as clusters (336 out of 355). The domination of clusters and standard interconnects will continue to exist in the future years, as this architecture provides great flexibility of scaling, from single nodes to large systems, and has superior price/performance over other architectures.
In order to understand the mixture of HPC and EDC systems in the cluster segment, one needs to analyze the application usage as reported on the TOP500 list. Since in several cases the applications are not known, and in other cases they are not clearly described, a conservative approach is preferred. All educational, research and classified applications should be marked as HPC, even though in some cases those systems are EDC type. Out of the cluster segment, 40 percent of the systems can be marked as HPC, meaning being used for HPC types of applications, while 60 percent of the systems can be marked as EDC.
The HPC part of the cluster segment is divided almost evenly between InfiniBand and GigE (45 percent and 55 percent). InfiniBand shows a steady growth in the HPC part, and is being driven by other technology trends as well – multi-core CPUs and fabric consolidation.
For the systems marked as EDC in the clusters segment, GigE connectivity is used in 85 percent of them and InfiniBand in 15 percent of the systems. Many of the EDC systems are supercomputers-for-a-day, and the rest of the time are being used as separate workstations, and therefore do not require high-speed connectivity. The average efficiency (how much of the theoretical available compute power can practically be utilized) of the GigE-based EDC systems is in the range of 50 percent versus 70 percent for the InfiniBand-based systems as reported on the TOP500 list.
This is yet another indicator that these systems are being used as small clusters or individual workstations. Low efficiency is translated into a vast waste of compute power and expensive maintenance (low price/performance and power/performance figures) and cannot be tolerated in supercomputers. Moreover, the EDC segment is traditionally known as a slow technology adopter, and therefore slower penetration of InfiniBand compared to the HPC segment is understandable.
Out of the InfiniBand-based solutions, 70 percent can be labeled as HPC and 30 percent as EDC, while out of the GigE-based solutions the case is opposite – 30 percent can be marked as HPC and 70 percent as EDC. While HPC and EDC share several key technology trends such as clusters and multi-core CPUs, some technology trends are applicable to only one of the markets. This is the case with virtualization. Virtualization aims to improve system utilization by enabling multiple applications to run in parallel on the same physical system.
Most of the HPC applications already fully utilize the system’s compute resources and therefore there is no need for virtualization. In EDC, the case is different and virtualization has become a key technology trend. By increasing system utilization through virtualization, the system I/O demands increase as well, to match the demands seen in the HPC environments. This leads to the increasing use of interconnects that can provide high bandwidth, instead of multiple GigE and Fibre-channel I/O ports (I/O unification). Therefore, one would expect to see increased use of InfiniBand for EDC systems in future TOP500 lists.
As in the TOP100, there is no 10GigE-based solution in the cluster segment, and the reasons are the same. 10GigE falls behind InfiniBand and GigE on the price/performance criteria, and with the adoption of virtualization, 10GigE will not provide the needed throughput and will fall behind InfiniBand. InfiniBand 40Gb/s has already been demonstrated, and with the InfiniBand technology being adopted in the EDC market, InfiniBand will become key for both clustering and virtualized environments. As power becomes an important factor in system design, power/performance becomes a key metric along with price/performance, and the ability to provide high throughput with low power will drive InfiniBand adoption as well.
The TOP500 list aims to rank the 500 highest performing supercomputers in the world, but as we enter a consolidation period, the list shifts from focusing only on HPC systems to representing systems used for other purposes as well, specifically EDC. Clusters derived from standard components have become commonly used for all types and classes of systems – from large-scale ten thousand server node systems, to small-scale systems with tens of servers, to single workstations. The small-scale systems and the single workstations or servers can be connected together to form a supercomputer-for-a-day, and be listed on the prestigious TOP500 list, and this has driven many OEM vendors to submit non-high-performance computing systems to the TOP500 list.
In order to analyze the vital information in the recent TOP500 list to better understand the technology and market trends, the list should to be divided into HPC-centric (TOP100) and general cluster (TOP101-500) categories. The former shows the trends in the HPC market and the latter in the general clustering markets, both HPC and EDC.
While the usage model of HPC and EDC systems and applications is different, most of the technology trends are identical. The need for complex simulations and research in the HPC segment and virtualization in the EDC market, together with the domination of multi-core CPUs and the need for faster storage, mandates the use of a high throughput and low latency I/O solution. InfiniBand is the only industry-standard interconnect that provides the required bandwidth, latency, power and utilization characteristics. InfiniBand has become the interconnect of choice for high-performance applications and has started to penetrate into the enterprise datacenter as well. With its superior price/performance and power/performance, InfiniBand is expected to continue those trends in the foreseeable future.
The author would like to thank Bill Lee and Sujal Das from Mellanox Technologies for their contributions during reviews of this article.
Gilad Shainer ([email protected]) is a senior technical marketing manager at Mellanox Technologies (www.mellanox.com) focusing on high performance computing, high speed interconnects and performance characterization. He joined Mellanox Technologies in 2001 to develop Mellanox’s InfiniHost PCI-X Host Channel Adapter (HCA) device and later led the development of Mellanox’s InfiniHost III Ex PCI Express HCA device. Gilad Shainer holds MSc degree (2001, Cum Laude) and a BSc degree (1998, Cum Laude) in Electrical Engineering from the Technion Institute of Technology in Israel.