IBM today announced that it has created a new distributed computing architecture with a General Parallel File System technology that is twice as fast as existing clustered file systems and that provides management and advanced data-replication techniques.
Calling it the General Parallel File System-Shared Nothing Cluster (GPFS-SNC), IBM said the new architecture is designed to provide higher availability through advanced clustering technologies.
Prasenjit Sarkar, a master inventor in storage analytics and resiliency for IBM's research branch, said the system scales linearly, so that a file system with 40 nodes would have 12GB/sec. throughput, and a system with 400 nodes could achieve 120GB/sec. throughput.
"It's very cost-effective bandwidth. You get 1MB/sec. per dollar," Sarkar said. "If you try to replicate that with a [storage area network], it gets very costly."
Sarkar described the GPFS's "shared nothing" cluster technology as each node or standard x86 server having access to its own metadata, cache, the data storage and management tools, while also having access to every other node in the cluster at the same time through Gigabit Ethernet ports.
"What we have done, in contrast to the Google file system, which has a single domain node, is we've distributed every aspect of the file system, the metadata, the allocation, the lock management, the token management," he said. "Even if you take out a rack of servers [from the cluster], we'll still be able to continue to work."
By "sharing nothing," Sarkar said, new levels of availability, performance and scaling can be achieved with the clustered file system. Each node in the GPFS-SNC architecture is also self-sufficient. Tasks are divided up between these independent computers, and no one has to wait on another, Sarkar said.
The GPFS-SNC code also supports Posix, which enables a wide range of traditional applications to run on top of the file system, allowing both reads and writes to be performed.
"You can open a file, you can read a file, then you can append to the file and overwrite any section. With Google's Hadoop distributed file system, you cannot append to a file, you can't overwrite any sections, so you're very limited in what you can do," Sarkar said.
GPFS-SNC also supports the whole range of enterprise data storage features, such as snapshots, backup, archiving, information life-cycle management, data caching, WAN data replication, and management policies. The architecture has a single global domain namespace, allowing virtual machines to be moved between hypervisor nodes.
"So for example in our cluster, you can run Hadoop as well as a clustered DB2 or Oracle databases," Sarkar said. "This allows us to have a general-purpose file system that [can be used by] a wide range of users."
IBM would not say when the GPFS-SNC file system would make it out of the labs and into the marketplace, but Sarkar said that once it it's available, it will be targeted at three use cases: data warehousing, Hadoop MapReduce applications and cloud computing.
"The cloud may not be intuitive of a parallel architecture, but we have [many] virtual machines on each hypervisor node, and we have a lot of hypervisor nodes in parallel. Each virtual machine is accessing its own storage independently of every other virtual machine. So in effect you're getting a lot of parallel access to storage," Sarkar said.
IBM's current GPFS technology offering is the core technology for the company's high-performance computing systems, Information Archive, Scale-Out Network-Attached Storage (SONAS), and Smart Business Compute Cloud.
The GPFS-SNC technology's ability to run real-time Hadoop applications on a cluster won IBM a first-place award at the Supercomputing 2010 conference in New Orleans this week.