Liquid Computing Brings Telecom Technology to Traditional Computing

Every start-up company claims it has unique technology that will change the marketplace. But most of these companies quietly fade off into oblivion, never to achieve their goals. We at IDEAS have seen dozens of these startups come and go over the years. However, Liquid Computing’s LiquidIQ fabric computing server caught our attention as one idea that may actually end up changing the market.

Liquid Computing was started in 2003 by two industry veterans: CEO Brian Hurley, who spent over 20 years at telecom giant Nortel, and CTO Mike Kemp, who amassed over 30 years of experience building high-end servers for organizations such as Nortel and the U.S. Defense Advanced Research Projects Agency (DARPA). Together, these two with their team of engineers have built an entirely virtualized server with a high-bandwidth, low-latency backplane to connect the processors, memory, and I/O subsystems. Best of all, they can offer this product for about the same cost as traditional high-end servers. Throw in autonomic system management that is straight out of the telecom market and you have a very interesting and desirable server. We recently had the chance to spend an hour with Brian discussing his current product and a looking at what Liquid Computing may have in store for the future.

Today’s LiquidIQ server supports twenty processor blade modules and five I/O blades in a chassis, delivering aggregate I/O bandwidth of 200 Gbps per chassis (supporting Gigabit Ethernet, 10 Gigabit Ethernet, and Fibre Channel). Each processor blade supports four dual-core "Rev F" Opteron 800-series processors and up to 16 GB of memory per socket. Two of these chassis, with a combined total of 320 Opteron cores, can fit into a standard rack, and in the current release a dozen chassis can be lashed together to form a 960-socket computing behemoth if desired. There are a couple secrets that contribute to the performance and scalability of this server. One is the proprietary IQInterconnect midplane that features a low-latency, 100 GB/sec bandwidth between processor blades; the other is the use of AMD’s Direct Connect Architecture and Hyper-Transport technology. The server can either be deployed as an 8-core coherent memory SMP server (16 cores when the new Opteron Barcelona blades are available), or it can remain non-coherent for traditional message passing, similar to that used in Linux clustering. Because it is 100% virtualized, configuration changes can be made on the fly and resources can be dynamically allocated where needed.

As expected, the LiquidIQ server is garnering quite a bit of attention from the high performance computing market. Anytime you mix fast processors, a high bandwidth low-latency interconnect, and a good price, the HPC market immediately takes notice. Add to that a world record STREAM benchmark result (demonstrating high memory bandwidth) for four-socket servers and a victory over the Cray XT3 in the HPC Challenge (HPCC) benchmark set, and you have a very nice HPC server. But what about the commercial computing market? That’s where the real money is to be made. This server already contains many of the features commercial customers are seeking, such as virtualization of all resources, integrated high availability capabilities, automatic detection and recovery of failed components, and capacity upgrades that can be performed online. To be successful there, Liquid Computing will need to achieve scalability beyond the 16-core level or end up competing against commodity server vendors such as Dell. It will also need to ramp up its support infrastructure. Commercial customers need a higher level of support than HPC customers. No matter which direction Brian chooses to take the company, IDEAS feels this is one start-up that is worth watching closely.