Analysis

posted on 29 February 2008 03:10

Where on earth does the name Panasas come from? Mount Parnassus is a mountain in central Greece that is associated with poetry, music, and learning. Panasas, the name of a Linux high-performance computing (HPC) cluster storage company, has nothing much to do with it at all, apart from the similarity in the sound of the word. There is a Panassas in Augusta county, Virginia, but that also seems to have little relation to Panasas, the company as that is head-quartered in Fremont, California.

'Pan' means all. ASAS is an offshore structural finite element system with integrated modules to address the specific needs of offshore and marine engineers. Since Panasas develops products for HPC applications perhaps this is a hint where the name comes from?

HPC market
The HPC market is a growing one. According to IDC, factory revenue for the HPC market grew 8.8 percent quarter over quarter and a full 18 percent compared to the same period last year, to reach $3.0 billion in the third quarter (3Q07). The worldwide server market, of which HPC servers are a part, grew 0.5 percent to reach $13.1 billion in 3Q07.

Third-quarter 2007 shipments of processor packages in the HPC market totaled 1.3 million, up 26 percent from the second quarter. In 2006, HPC systems accounted for 26 percent of all processors sold in the server market, more than doubling the 2003 HPC share of about 12 percent.

(Panasas ActveStore cluster pictured.)

Linux clusters
What Panasas does is produce parallel access storage for clusters of Linux processors in HPC environments. It produces proprietary hardware and software in the form of a parallel access storage appliance. This runs PAN FS, an object-based, parallel file system, similar to Lustre and GPFS. It provides for direct and parallel access from each processing node in a Linux cluster to the storage. There is a single name space and near-linear scalability as nodes are added.

To combat any data corruption risk there is tiered parity with parity at the disk level up to the cluster level. Panasas also offers per-file RAID for more granular protection. This all allows for very fast file reading and writing and also fast reconstruction if a disk should fail.

Derek Burke, Panasas' commercial director for EMEA, says that 80 percent of HPC customers are using Linux clusters with commodity computing nodes. He says HPC customers need a parallel-access storage facility because processing nodes are being added faster to clusters than storage bandwidth can be increased. Nodes are being added all the time as data sets get larger and analysis runs make more and more iterations through the data.

I/O limitations
Software tools to parallel process the analysis on such data sets are getting better. But CPU speeds aren't increasing much. Cluster processing power is rising because more cores are being added to processors and more processing nodes are being added as well; a kind of double whammy.

The net effect is that more individual instances of an analytical app in the cluster need access to storage. A 100-node cluster changing to dual core nodes has a doubling of its basic set of application threads needing access. It's as if the storage access ports from the processing part of the cluster double yet the actual access ports into the storage part of the cluster stay the same. That's called a choke point in most people's view.

Burke said: "This means storage I/O to network-attached storage (NAS) and clustered NAS becomes a bottleneck because of its serial I/O. Put 32 nodes or more up and any NAS shows bottlenecks starting. There is downtime and this decreases cluster utilisation. The bigger the cluster the bigger the downtime problem."

"Clustered NAS addresses part of the issue; it can handle data quickly and efficiently but there is still no parallel access across all the nodes to the storage. There is still an I/O read/write wait time problem."

He talked about CFD - computational fluid dynamics - applications in which data is saved at multiple points in the run time period. Any I/O wait time just lengthens the job and reduces cluster productivity. The answer is to provide parallel access from all the nodes to the storage resource: "A storage limiter on HPC is removed."

Panasas architecture
Panasas has a bladed architecture and its storage nodes connect to cluster processing nodes across 1 or 10 gigabit E, meaning up to 400 or 700MB/sec bandwidth per shelf. Panasas blades can be either director blades or storage blades, which handle I/O and storage respectively. There can be up to three director blades per shelf and 8-10 storage blades per shelf. The director shelves can have up to four Ethernet connectons. The storage blades use either 750GB or 1TB SATA drives providing massive capacity across lots of spindles. There can be ten shelves in a full height rack and Panasas racks can be connected together.

Pan FS is object-based and allows per-file RAID. This is useful for very large files Burke says: "They are spread across disks, blades and shelves. Multiple entities rebuild their parts of a file (meaning) very fast reconstruction."

Panasas situation
The company is in a pre-IPO situation.

Burke says the HPC market is a growth one with lots of demand for Panasas' product. It is anticipating extremely strong global and European revenue growth this year. He said: "The vast majority of HPC customers are considering deploying parallel storage systems. We've been having lots of success. HPC is growing rapidly through commodity computer building blocks. There is strong European revenue growth." The company has fifteen people in Europe.

HPC is finding a role in life and material sciences, computer-aided engineering, oil and gas, visualisation, climate computing, weather and oceanographics, digital movie rendering and high-end business analysis in financial services. Panasas is: "more focussed on manufacturing, energy, higher education and government."

Does Panasas see Blue Arc in its sales encounters? "We don't see them much in HPC." The feeling coming through here is that neither HW-assisted NAS nor clustered NAS, think Isilon, can help sort out the HPC cluster I/O bottleneck as effectively as a parallel storage system, in Panasas' view.

What is Panasas about? It provides parallel-access and high-capacity storage for clusters of Linux computers in the HPC environment. With its storage HPC jobs run faster and cluster processors spend less time waiting for I/O to complete. That's the Panasas product story in a nutshell.