17 Comments

Summary:

Schooner Information Technology, a 2-year-old year old startup in Menlo Park, Calif., today came out of stealth mode with an appliance designed to speed up the transfer of information. As online data becomes more prevalent and the patience to wait for that data wanes, the company […]

schooner-appliance-4-8-09Schooner Information Technology, a 2-year-old year old startup in Menlo Park, Calif., today came out of stealth mode with an appliance designed to speed up the transfer of information. As online data becomes more prevalent and the patience to wait for that data wanes, the company is offering a machine that’s purpose-built to speed up memcached and MySQL for web-scale applications. This means faster-loading photos for a site like Facebook and faster access to databases for serving up other information online.

Schooner has created specialized machines optimized to run those software applications at more rapid speeds than a commodity server would. The payoff is better performance, a 60 percent reduction in energy use, and a space-saving box. The company, which has raised $15 million from Redpoint Ventures and CMEA, has built a whopper of a machine that packs the latest Intel Nehalem server chips, 512 GB of memory, and 1-Gigabit Ethernet or 10-Gigabit Ethernet in a 2u rack.

The hope is that the Facebooks and Googles of the world will choose a single Schooner machine, which costs $45,000, in place of eight or more commodity servers running MySQL or memcached. Those commodity servers can start at $600 apiece and increase in price  depending on how they’re configured.

Schooner’s underlying premise in offering the appliances — essentially that hardware needs to be specially tuned for software — is one that’s being explored in greater detail by vendors ranging from Microsoft to VMware. However, I wonder if energy savings and better performance are enough to lure giant data center operators away from cheap commodity servers.

Schooner’s co-founder and CEO, John R. Busch, says that some smaller web operations and those overseas would rather buy purpose-built machines instead of building custom software to improve performance on commodity boxes. Many big-name web site operators, such as Amazon or Google, roll their own software, but Schooner thinks there’s a market in folks who operate web-scale data centers and just want to buy an expensive box and forget about it. Busch may be right, but I’ve seen appliance efforts play out before. And I can’t help but think of that abandoned panini press buried in the back of one of my cabinets.

  1. First is was XML and Appplication router appliances that were the biggest losers ever birthed by an industry. DataPower and Reactivity come to mind. The problem: the design cycle for specialized logic and custom accelerator ASICS is measured in years, while the iterations for increases in general purpose server architecture is a match.

    while the XML box people were banging the drum (oh, you fools), multicore chips were just getting into low cost servers – they must have had that sinking feeling. None of those XML and App router companies ever got big, and most who were acquired by bigcos got lost in the maelstrom.

    On the other hand, general purpose server images with Kernel enhancements for application specific optimizations, now you are talking.

    Share
  2. [...] Schooner Launches Specialized Servers for Speedy Data Delivery: “GigaOM” [...]

    Share
  3. Bret Clement Monday, April 13, 2009

    Gear6 will officially announce their new memcached-based appliance on April 20. This product recently came out of beta. Learn about it on this April 14 webinar: http://bit.ly/cvFWy “Scaling website performance by caching session and profile data with Memcached” or read about it in today’s press release: http://bit.ly/85hYH.

    Share
  4. One more thing:

    I’m not banging these guys at Schooner, specifically, all power to them – as if my opinion matters at all. The general trend has been for application specific boxes, like XML accelerators, to just never get traction in the market.

    That said, there is a problem with today’s n-tier model, not actually the model, but the way it has been pushed into the field. That connection between application server / logic, and database is just such a mess, and it has spawned these bastardized tool-sets to cure the illness. I’m just not sure that custom boxes are the way to address this crappy coupling of database to app server.

    See Openlinksw.com, where they solved the issue of coupling years ago. They got no respect for the brilliant and elegant solution of vertically integrating the app server, language hosting run time environment, and the database server. As a company, they have solved so many legacy problems and moved on, now to the CEO’s evangelism of Linked data, which I completely fail to comprehend; Kingsley lost me two years ago at RDFS.

    Share
  5. What does it cost? Has it done anything for enhancing I/O bandwidth, minimizing seek times etc? it seems optimized for read-mostly web applications (big ram [for memcached?], large network bw support).

    Share
  6. I must confess that I am having a hard time working out a viable economic scenario for these appliances.

    An observation about appliances generally: if you are big enough to need extreme performance that might justify buying an expensive appliance rather than using a large, generic server, you usually have someone around who can tweak and tune the server application to get a similar result less expensively. It is a rare company that has Internet-scale web apps with racks of servers and yet lacks the resources to administrate them such that it would justify an appliance as a shortcut. For certain very specialized domains there might be an argument for it, but not for generic applications like MySQL and memcached.

    Specialized appliances find their sweet spot in the market as an abstraction layer for complex functionality that fits poorly in vanilla server hardware (think core routers) or by using the appliance as an interface to unique functionality that offers more than just first-order convenience (e.g. some database analytics appliances). Simple convenience as the abstract selling point for an appliance is a hard position to defend over any length of time; someone will inevitably separate the convenience from the hardware, and sell it at a significantly lower price.

    Share
    1. How is the ‘Schooner Appliance’ any different than a large generic server with a new name? It looks to me like it’s just a server with a lot of RAM and some solid state storage.

      Share
  7. Google / Amazon / Yahoo / MSFT et al actually see $600-1000 servers as an advantage.

    You get more of them, and if they fail, they can be:

    a) replaced cheaply
    b) take down only 1/x of the infrastructure dedicated to that service/pool

    Putting your eggs in a giant basket/schooner is precisely what they don’t want to do @ their scale. Data center engineers @ GAYM assuming failure as a constant in their environments. Distributing that failure equates to lowering the probability of sustained downtime and impact.

    Additionally, the granularity allows a data-center to evolve to more modern hardware smoothly… decommissioning a machine doesn’t mean loss of service.

    Enterprises are a different story, and schooner may have a shot there.

    Share
  8. I totally agree with zubinwadia. Unless its going to show some reeeeaaal significant improvement otherwise I don’t think the current loading speeds are any major problems.
    I guess the advantages just doesn’t weigh out the disadvantages.

    Share
  9. Memcached (www.danga.com/memcached/) is a trivial less than 2000 line of code. What effort does it take to have it run on a $400 SSD? Why not start an open-source project on it and take Facebook’s network and scalability optimizations and build free software for everybody?

    Share
  10. I run 14 racks in a data-center for a Web 2.0 web-site (cannot share the name) on LAMP stack.
    I am not sure if there are just developers making comments here. Developers dont have to pay for their developed application, I do. I run a script that collects “unix load” and application requests per second across the servers. Let me make some quick observations for many unaware developers::
    1) There is no $600 server. If there is no memory and no storage throughput on the system, its not a memcache or mysql box. It could be a webserver, but not memcache or mysql – lets not confuse.
    Our servers cost ~$6k for memcache and $10-12k for mysql servers.
    2) 97% of the time, the load is below minimal threshold set by our sizing team – that adjusts load and sizing on memcache and mysql every month.
    3) The cost (power + cooling) of 14 burning racks is astronomical – I wish we were getting paid for every transaction like EBay. I cannot even depreciate these costs.
    4) So one year ago, someone said Virtualization, now we spend 1 person on provisioning servers and taking care of images. I never knew virtualization would need so much administration. Solutions offered to me are hefty in their price or devoid of any good functionality set.
    5) 6 months ago our lead systems engineer took on the challenge of coding up memcache on Flash. After going through various vendors and different technologies, $100k down the drain, performance, stability and functional objectives havent met. That’s not to say other smart engineers wouldn’t be able to crack the nut… but I would rather have had the systems engineer work on problems that our users hit, which relieves immediate pain.
    6) The $400 SSD is probably MLC which suck as you cannot do a lot of writes to them.

    Bottom line, I like to see real hard data, not marketing hype. Bring me an appliance that solves my problems, I will negotiate the price and get a better bottom line for my company. Now thats only what I care for.

    Share
    1. Aaron deMello Monday, April 20, 2009

      Current gen MLCs can support a large number of writes. Check out the recent reviews on Anandtech and Tom’s Hardware.

      If you blew $100k on your SSD / Memcache project and have nothing to show for it, then there is something wrong with your engineer or perhaps you have a rare app that does not benefit from cache…. which I find hard to believe.

      I agree that you power costs must be astronomical… this would be helped tremendously if you built a cache layer that was hit before your dbase layer.

      Finally, you can get a 32GB SSD that is SLC for $400.

      Share

Comments have been disabled for this post