3dmentat - Fotolia
A3Cube Inc. this week fleshed out its strategy for boosting the performance of I/O-intensive applications with the launch of its software/reference architecture designed to aggregate and scale memory and storage resources.
The startup's newly released Fortissimo Foundation complements the PCI Express-based Ronniee Express network interface card (NIC) made available in February. A3Cube claims a storage server cluster running the Fortissimo software and Ronniee NICs can significantly improve the performance of NoSQL and legacy SQL databases and Hadoop-based analytics, scientific, multimedia and virtualized applications that run across multiple servers.
Emilio Billi, CTO and founder of A3Cube, said existing scale-out storage approaches often suffer from bottlenecks driven by the high latency of the fabric used to connect the scale-out nodes. They also require a metadata server that is a single point of failure for the system and a single point of congestion for the I/O-intensive applications.
Fortissimo Foundation virtualizes a standard file system across a cluster of commodity x86 servers and provides a single global namespace. In combination with the Ronniee NIC, the software aggregates and manages the cluster's memory, disk and bandwidth resources. That makes them appear local, and distributes processing power across the cluster, according to Billi.
"We provide parallel access to all the resources," he said. "We don't need an API like in object storage to be able to read and write into the storage pool. Everything happens in the same way it happens in a local machine, with bare-metal performance."
Billi compared A3Cube's conceptual approach to object storage running on a multimillion-dollar, high-end symmetric multiprocessing system. He said the A3Cube system avoids the performance penalty associated with the object storage API and provides direct access to the merged resources of potentially thousands of nodes at a fraction of the cost.
The Fortissimo Foundation software stack runs on standard Linux-based servers equipped with any type of storage media, including solid-state drives (SSDs), hard-disk drives (HDDs) and flash dual in-line memory modules. Fortissimo sells for $5,000 to $6,000 per node, regardless of capacity, and the accompanying Ronniee Express NIC costs under $1,000.
The Fortissimo software must be installed on each server node in the cluster. The default file system is ZFS, although Billi said Fortissimo supports any POSIX-compliant file system, including GPFS and ext4. The software also supports storage features such as replication, automated tiering, and inline data deduplication and compression. A3Cube is also working on a plug-in for geographic replication, according to Billi.
The Ronniee Express NIC plugs into a PCI Express slot on each node and provides low-latency, in-memory communication between remote and local servers. It bypasses the operating system kernel to accelerate the resource aggregation, according to Billi.
"From a system point of view, this is NAS connected to the system data center using NFS, CIFS, HTTP, iSCSI or whatever, using the existing data center interconnection fabric," Billi said.
Throughput, IOPS and capacity scale linearly with the addition of server nodes. The maximum tested configuration was 120 nodes, but Billi said the fabric can theoretically scale as high as 64,000 nodes and the file system can potentially reach 16 exabytes.
Billi said the A3Cube-based system makes no distinction between memory and storage. Users can configure storage tiers and designate DRAM as the fastest tier and slower SSDs and HDDs as lower tiers of storage, if they choose.
Customers can connect and aggregate legacy storage, and they also have the option to consolidate compute, virtualization and storage into an A3Cube-based hyper-converged clustered system. A3Cube currently supports the open source KVM hypervisor and is working to add support for VMware, according to Billi.
Use cases include HDFS, MySQL
A3Cube can also enhance Hadoop Distributed File System (HDFS) and MySQL. Customers can virtualize the HDFS on top of an A3Cube cluster to enable Hadoop to use HDFS at the speed of DRAM. They can also run a standard MySQL database in an A3Cube-based DRAM cache to emulate the performance of HANA without having to buy special software or extra hardware, Billi said.
"What's really unique is that they use this Ronniee card that's plugged into each server and mini-SAS cables to connect up to the various servers in the cluster," said Jim Miller, a senior analyst at Enterprise Management Associates, based in Boulder, Colo. "The difference between a traditional virtual or software-defined approach is that the latency is really, really low because they don't have to go out over Ethernet or InfiniBand to get to those other resources. They can use this really fast server-to-server connection via PCIe. In the end, scale-out applications are able to just run a lot faster with fewer resources because it's all shared across all the servers."
Miller said that while the potential is great to leverage resources and increase scale-out application performance, the "proof is in the pudding," and he looks forward to seeing real-world results.
A3Cube has lined up four system integrators -- Penguin Computing and three others that it declined to name until they finalize product testing. An A3Cube spokesman said many customers are testing the product, but to date, only one has purchased it.
Scott Sinclair, an analyst at Enterprise Strategy Group, said that as organizations build out their infrastructure for big data analytics, they will tend to stick with their existing infrastructure without investigating alternatives. "A3Cube's challenge will be to demonstrate that their technology provides enough benefit for organizations to try something new," Sinclair wrote in an email.
"Despite all the noise around big data, the market is still in its infancy," Sinclair added. "That can be both a risk and opportunity. A3Cube has some interesting technology and some customers signed on, but it remains to be seen how widely applicable their technology will be. It is just too early to tell."