Original URL: http://www.theregister.co.uk/2009/04/13/schooner_memcached_mysql/

Big-iron brains powers Schooner appliance power

Putting a ding in server size

By Timothy Prickett Morgan

Posted in Servers, 13th April 2009 20:53 GMT

A start up driven by some big-iron brains has emerged from stealth mode to deliver, no, not a better server but a duo of server appliances - one to accelerate web caching and the other to speed up MySQL databases.

Schooner Information Technology was founded in February 2007 by president and chief executive officer John Busch, and chairman and chief technology officer Tom McWilliams.

Both have very long and deep experience in IT, and given the enthusiasm for server appliances and any technology that will help data centers save money and speed up performance, Busch and McWilliams had no trouble raising cash. They got $15m in Series A venture capital from CMEA Capital and Redpoint Ventures to fund the research and development behind their appliances, which already have 11 patents filed concerning their design.

So what of those backgrounds? Busch worked his way up through the ranks of Hewlett-Packard, and for a dozen years was in charge of the technology behind its HP 3000 proprietary minicomputer platform - the one that predates the company's entry into Unix by a decade. He was also co-founder and head of engineering at Clarity Software, and after that worked as the vice president of engineering at a network appliance maker called Diba, which Sun Microsystems acquired in 1997.

Once at Sun, Busch headed computer systems research at Sun Labs, doing some of the original research that lead to the chip multithreading architectures on the Niagara Sparc T and Rock UltraSparc-RK that Sun hopes will preserve and extend the Sparc's reach in the enterprise.

McWilliams was a supercomputer designer at Lawrence Livermore National Laboratory, and then in the early 1980s, he founded a company called Valid Logic Systems, which did a lot of the basic research behind computer aided-engineering systems. In 1987, McWilliams founded a server maker called Key Computer, which made one of the first superscalar pipelined processors and which was acquired by clone mainframe maker Amdahl in 1987. That left McWilliams in charge of Amdahl's systems architecture group.

In 1993, McWilliams jumped to Silicon Graphics and took over development of its MIPS RISC chips, which he did until Sun bought the Starfire server line from SGI. From 1996 to 2001, McWilliams work on Sun's server architecture and also on CAD tools. After leaving Sun in 2001, he founded PathScale, which created a supercomputer interconnect based on InfiniBand, a company that was acquired by QLogic in 2006.

You'd think these two would get together and build a better server. And in a way, that is what they have done. But rather than build general purpose machines, they are instead creating appliances tuned to do specific jobs and that help knock out large numbers of general purpose machines in the data center that don't do these jobs very well.

"We took a hard look at scale-out architectures - Web 2.0, cloud computing, and the like - and what we saw was a high amount of power consumption, complex software, and a lot of inefficient iron," Busch said. "We saw a large opportunity to make a dent in this."

That's what a lot of vendors and IT managers see, but some IT vendors - those that make their dough selling x64 iron or operating systems and middleware stacks running on them - actually make their money because of the inefficiencies, however much they bemoan them.

The first workload that Schooner is attacking is web caching, something that is happening every day behind all of the web pages we surf when we are supposed to be working. To make a web site hum - where you want to get complex queries on search engines to come back in a hundred milliseconds or so and for those queries to hit tens or hundreds of servers to get their results - you have to cache everything, not just the results of queries.

And so, Web 2.0 companies have racks and racks of x64 servers sitting there caching things, usually employing an open-source program called Memcached.

Schooner thinks it has built a better Memcached mousetrap. It has taken one of IBM's two-socket System x servers using Intel's new Nehalem EP Xeon 5560 processors - four-core, eight threads, running at 2.8GHz - and then stacked eight of Intel's 64GB X25-E Extreme solid state disks - which are arranged vertically inside the System x's 2U chassis - for a total of 512GB of flash memory; the machine also has 64GB of DDR3 main memory.

The server can have Gigabit Ethernet or 10 Gigabit Ethernet links. That is the only configuration. The appliance runs a standard Linux distribution, and it is unmodified. The company did not say which one.

There's nothing too special about this hardware or using Linux - anyone can do this. Where Schooner has done its R&D is in creating a completely black box - and not open sourced - clone of Memcached that has been reworked to take advantage of the multiple cores and instruction threads in modern processors like the Xeon 5500s and it is also keenly away of memory, both flash and main, inside the box.

Just slapping the open source Memcached on a Nehalem EP box is not going to yield the performance benefits that companies expect, Busch said. "The industry has gotten the pieces optimized locally," he explained in reference to server makers using better chips, more main memory, and now integrating flash. "But the hardware and software is not being optimized to run together."

Schooner's own Memcached variant supports the same APIs as the open-source Memcached, so applications won't see any difference. And to steer clear of open source licensing issues with Memcached or Linux, Schooner's Data Fabric API, which is an operating environment that takes over the cores and their threads plus main and flash memory access and scheduling, as well as the Memcached program runs in a userspace inside Linux. There are no mods to the Linux kernel.

Based on its own benchmark tests, the Schooner Memcached appliance is able to make use of all 16 threads in the box, and can handle 300,000 Memcached requests per second, compared to just under 50,000 requests per second with the open source version of Memcached on the same Nehalem iron. And if you compare the Schooner appliance to a Memcached setup running on a prior generation of Intel Xeon iron, the difference is more like a factor of 8X in performance.

Even at $45,000 a pop for the Schooner Memcached appliance, the savings over plain vanilla x64 servers can be dramatic, according to the company. As an example, Schooner has calculated that it would take 167 x64 servers to deal with a mid-sized Web 2.0 data center with 5TB of data it needs to cache. It would take about $1m to buy the iron, and a little more than $2m more to keep those x64 servers maintained over the course of three years, including power and cooling.

Installing 20 Schooner Memcached appliances will cost about $900,000, but maintaining, powering, and cooling these boxes will only cost an additional $554,000 over three years, more than cutting down the cost of supporting the same Memcached workload.

The second appliance that Schooner have launched has the same physical hardware, but uses a highly optimized version of the InnoDB 1.0.3 transactional storage engine for the MySQL database, which is controlled by Sun. The appliance also includes a copy of the MySQL 5.1 Enterprise Edition database, which has been OEMed from Sun. By switching to flash-based SSDs for database storage, the Schooner MySQL appliance has demonstrated that it can do about eight times the online transaction processing throughout on the TPC-C test compared to last year's pre-Nehalem Xeon DP server.

In a similar Web 2.0 data center with that same 5TB database, Schooner figures it would take 102 database servers running MySQL to so the job, and over three years, it would cost $2.55m to acquire, power, cool, and maintain those boxes - a little less than half of this is capital expenses to buy the iron.

But using its MySQL appliance, which will also sell for $45,000 a pop, Schooner thinks this data center will be able to get by with 13 appliances at a cost of $585,000 and then spend another $371,770 to maintain, power, and cool them over three years. This chops out nearly $1.6m from the IT budget to support the same workload. Presuming, of course, the hypothetical Web 2.0 company is buying all new gear.

Still, by Schooner's numbers, the operating expense savings from using the Memcached and MySQL appliances and maintaining them covers a lot of (but not all) of the cost of using these appliances. In tightly constrained data centers where power is an issue, this will be enough to open the door to negotiations.

The two appliances will be available in May, and were co-developed with IBM's Systems and Technology Group. It is not yet clear if Big Blue will be a Schooner reseller, but it seems likely. ®