As the economy rebounds from the recession, high performance computing seems to be recovering in kind. IDC says HPC revenue grew by 2 percent in the first half of 2010 and predicts the market will grow by 6 to 7 percent for the year. That rosy picture was reflected in some notable HPC success stories over the last 12 months. But there were a few setbacks as well. Here are my hits and misses for the year.
Hit: GPU Computing Comes of Age
No surprise here. Led principally by NVIDIA, GPU computed made some significant headway in 2010. Nearly every HPC system vendor added Fermi GPU-equipped computers to their stable this year, including the big three server makers: IBM, HP and Dell. On the software side, CUDA applications and third-party integrations proliferated, while ISVs like ANSYS, SIMULIA and Livermore Software Technology Corp. (LSTC) announced plans for GPGPU support. Oh, and as of November, three of the top four supercomputers in the world are powered with GPUs, including the number one system from China.
Hit: China Takes Supercomputing Crown from US
Speaking of which: Yes indeed, the Asian superpower gave birth to the number one supercomputer in 2010. Tianhe-1A delivered 2.56 Linpack petaflops, knocking ORNL’s Jaguar (and the US) into the number two position. China also took third place with Nebulae, a 1.27 petaflop system. The country now has 24 systems on the TOP500 list, and given its stated ambitions, intends to grow that significantly in the coming years.
Miss: Oracle Shows HPC the Door
After aquiring Sun Microsystems in January, Oracle has explicitly ignored its newly acquired high performance computing assets. Certainly Sun’s mainstream HPC blade and storage lines have been set adrift. The company is still the caretaker of Lustre, the open source parallel file system it inherited from Sun, but all internal development will be targeted to Solaris and its own database machinery. Oracle’s motivation to buy a 10 percent share of Mellanox appeared to be along the same lines — to keep its database business fed with InfiniBand technology. For the foreseeable future, the company seems content to be a user of HPC, rather than a vendor.
Hit: Lustre HPC Community Reorganizes
Oracle’s abandonment of Lustre for HPC has managed to rally the rest of the community. In July, a company called Whamcloud emerged, whose mission is to take up the HPC-Lustre cause. Then in October, a non-profit named Open Scalable File Systems, Inc. (OpenSFS) was formed to bring together all the stakeholders. And just this week, a Europe-based group was organized much along the same lines. For the time being at least, Lustre’s HPC future seems safe.
Hit and Miss: HPC On Demand
OK, it looks like there could be something to this whole “cloud” thing all the kids are talking about. From consumer-side apps to big mainstream enterprise services, cloud computing has made some enormous strides this year. As usual, in HPC things are a bit different. At this point, HPC users are mostly just kicking the tires. Use cases are rare. The big breakthrough this year: Amazon launched an HPC instance for its EC2 computing on-demand service (sort of the Good Housekeeping Seal of Approval for HPC in the cloud). The longer view says cloud computing is destined to be a driving force in HPC, just like ummm… pNFS.
Miss: pNFS Misses Its Mark
Parallel NFS (pNFS), the standard that will provide scalable parallel file access for NFS users, has been hashed over since 2003, and is just now getting ready for its big debut… again. The technology was supposed to deploy in late 2009, and then 2010. We’re still waiting. The HPC storage vendors, especially Panasas and BlueArc, along with other interested parties like NetApp and Microsoft have been generally cheerleading the effort, but as with all standards that push a large community forward, timelines often get stretched. Now the expectation is that we’ll see the first commercial pNFS solutions sometime in 2011.
Hit: SMP Resurgence
Single image, shared memory computing has always provided an enticing alternative to the messiness of distributed memory, cluster computing. With the proliferation of cores (8 or more in high-end CPUs these days), increasing memory capacities, and more performant interconnects, SMP-style high performance computing is now more cost-effective than ever. The SGI Altix UV super, announced at SC09, and which began shipping this year, is the most notable example. But there are plenty of other alternatives out there, including newer products like Numascale’s NumaConnect SMP adapter, bullx supernodes, the Cray CX1000-S machine, as well as virtual offerings from the likes of ScaleMP, RNA Networks, and Symmetric Computing. With yet more cores and even faster interconnects on the horizon, the SMP trend can only continue.
Hit: InfiniBand Surges
Despite the industry push behind 10 Gigabit Ethernet, in HPC InfiniBand use continues to grow and in some cases, dominate. On the latest TOP500 list, the InfiniBand claims 215 systems while Ethernet-connected systems are down to 227, representing an 18 percent increase for the former and a 14 decrease for the latter compared to 2009. And for less elite systems, InfiniBand appears to be the interconnect of choice. With the roadmap headed torward 56 Gbps (4X FDR) in 2011 and 104 Gbps (4X EDR) in 2012, InfiniBand’s momentum seems assured.
Hit: Supercomputing for Everyone
The petaflop milestone seems to have created a rush of entrants into the elite segment of HPC. All of the top ten systems in the world are now (peak) petaflop machines, and half of them are housed in countries outside the US. As mentioned before, China is making a concerted investment in big machines, but Europe is also methodically building its petascale cred. And the supercomputing end of market continues to be among the most robust. Even in the down year of HPC, high-end revenue increased by 65 percent, although growth is expected to be flat for 2010 (this according to IDC). In any case, big machines are increasingly seen by nations, states, and organizations as the catalyst to cutting-edge science, so I expect these elite machines to continue to be hot commodities.
Hit: Olympic Bobsledding, HPC-style
The feel-good HPC story for 2010 has to be the success of the US Olympic Bobsledding team. The four-man team was able to capture not just the gold medal, but the course record as well. The HPC connection is that the bobsled was designed by CFD software from Exa Corporation, using a modest-sized x86 cluster. Through the CFD simulations, they were able to improve sled aerodynamics by around 2 percent, which turned out to be the margin of victory. And you thought HPC was only good for depressing stories about global warming.