But Wait, There's More

Cray Has Coming Out Party for XD1 Linux Supers

Supercomputer maker Cray last week had a coming out party for its Linux-Opteron based XD1 supercomputers. The XD1s are, like the "Strider" variants of the Red Storm supercomputer that Cray is building for Sandia National Laboratories, based on Opteron processors and run Linux. While the Red Storm machines are designed to scale to tens of thousands of processors, the XD1s scale well from a few dozen Opterons (with special assist FPGA logic) to thousands of processors. Cray has put the XD1s through their paces at Ford Motor Company, the U.S. Army, and Boeing (to name a few companies) and is now offering these supers for sale. (See "OctigaBay Takes Opteron-Linux to New HPC Heights" for a full description of the XD1 machines. The main difference between that description and the machines that are shipping today are that the XD1s support faster Opteron processors.)

While the underlying XD1 technology may not have changed since OctigaBay launched its machines in February and Cray snapped it up for $115 million before it could get rolling on its own, the benchmark tests that prove the XD1s can do what OctigaBay said it could are the real news this week. Parallel supercomputers work by crunching numbers and passing data (called messages) between node in the clusters so one node that has processing that is dependent on the results in the other node can do its work. The speed of the Message Passing Interface (MPI) protocol is, in many ways, more important than the aggregate CPU power for many parallel workloads. Adam Lorant, vice president of marketing at Cray, says that many of the servers used in commercial two-way and four-way servers were "optimized for TPC, not for HPC," which means that they have processors operating asynchronously on systems passing very small messages from many clients through transactional systems. HPC systems may have only one use, and the processors have to work in concert on a job that may run for weeks or months. Which is why Cray is pushing the "purpose-built" XD1 in a supercomputing world increasingly dominated by Lintel and Lopteron clusters. Lorant says that the RapidArray interconnect embodied in the XD1s has a latency that is one-fourth that of InfiniBand interconnect and that it offers twice the message throughput for 1 KB messages. To put it into perspective, an XD1 will offer about 30 times the I/O bandwidth of a Linux-Opteron cluster using Gigabit Ethernet to link the machines to each other. On one weather modeling application called MM5, the XD1 yielded 40 percent better price/performance than a Xeon/Myrinet cluster with the same processor count.

A 12-way XD1 bay with custom electronics for interconnect costs about $50,000 with Opteron 248 processors, 2 GB of main memory, six disk drives, 4 GB/sec of interconnection between nodes, and a dual Gigabit Ethernet switch for talking to the outside world. Cray expects its typical XD1 customers to buy from 2 to 40 of these 12-way chassis, spending upwards of $2 million. The XD1s run a modified version of Novell's SuSE Linux Enterprise Server 8, and will soon move to SuSE Linux Enterprise Server 9. One of the Linux tweaks that Cray created for the XD1 was to reduce "OS jitter," where processors can get out of phase doing housekeeping functions and therefore be unavailable for actual HPC work. By synchronizing the housekeeping functions on all processors in the XD1, Cray has been able to boost the performance of the cluster. By launching the X1 parallel vector servers last year, Cray moved from 4 percent of the $5 billion HPC server market in 2002 to a 20 percent share of a $6 billion market in 2003, and it is hoping that it can get a big bump in market share in 2005 as the XD1s are accepted for customers who want Linux and Opteron clusters, but who want to run more efficient systems.

HP Promises Specialized Linux Telecom Blade Servers

Hewlett-Packard announced last week that it would create a line of specialized blade servers for the telecom and service provider industry that meets the AdvancedTCA standard for telecom blades. This AdvancedTCA standard is based on a prevailing telecom blade server standard that predates the idea of commercial blade servers from the late 1990s and has been used for about a decade. Telecom equipment makers and telecom companies that used to build their own gear have long since used blades, but they need designs that have a window of usage that spans a decade or more.

To this end, and to better compete against rival Sun Microsystems, which gets a large portion of its server volume selling into the telecom market, HP has decided to build X86-based AdvancedTCA servers that support Carrier-Grade Linux and thereby chase some of the telco business that Sun has locked in with its baby Sparc/Solaris boxes. Joy King, director of HP's network and service provider solutions group, says that the company will launch the new servers as the Advanced Open Telecom Platform, which will include NEBS-compliant commercial-grade and carrier-grade servers based on 32-bit Xeons (called the cc2300s), 64-bit Xeons (called the cc3310s), and 64-bit Itaniums (called the Integrity cx2600. Linux will be the platform for these boxes, but where telecom companies require it, it will support Windows or HP-UX (on the Itaniums only for the latter). HP is also planning to push its NonStop fault-tolerant Unix boxes as back-end systems for this switching gear. The AdvancedTCA blade servers will also include specialized storage and switching blades. The first such machines will ship sometime in 2005.

In December 2003, HP inked a deal with RadiSys to incorporate that company's PICMG and AdvancedTCA blade servers into HP's carrier-grade server portfolio. It was unclear at press time if these RadiSys machines were what HP was announcing, or if it had decided to build its own boxes, or use a mix of HP and RadiSys gear.

OSI Hosting Buys 12,000 SuSE Linux Enterprise Server Licenses

Dallas-based OSI Hosting, a service provider established in February, and commercial Linux distributor Novell made an unusual announcement last week and said that the company was deploying 12,000 SuSE Linux Enterprise Server licenses, worth some $105 million, just like it said it would back in June, and said further that it was not in talks with alternative Linux suppliers for its hosting operations. This is arguably one of the largest Linux contracts to date, which is why tongues have been wagging.

OSI Hosting is running SuSE Linux on IBM's xSeries servers, and is selling dedicated hosting for as little as $80 a month on Xeon and Opteron boxes. The company has also built another setup based on Apple Xserve PowerPC machines and Mac OS X, but hosting on these boxes costs $299 a month.

Novell Releases SuSE Linux Professional 9.2

Having just revamped its desktop Linux earlier this year, Novell is keeping the heat on rival Red Hat, which has renewed aspirations on the desktop, by shipping its SuSE Linux Professional 9.2 release in November.

SuSE Linux Professional 9.1 debuted in March, and it was Novell's first stab at using the then-new Linux 2.6 kernel. With the Professional 9.2 release, SuSE is going to make it easier to set up connections to Bluetooth devices and wireless LANs by improving its YaST system management program, one of the key differentiating features of SuSE Linux over the years from alternative distros. Professional 9.2 will also support the 64-bit memory extensions in the AMD Opteron and Intel Xeon processors, which are making their way into PCs and workstations these days. The release will support the KDE 3.3 and Gnome 2.6 graphical user interfaces, the OpenOffice 1.1.3 office suite, the Evolution 2.0 email and calendar client, the GIMP 2 graphics tool, and the Nvu Web page authoring tool. SuSE Linux Professional will be available in early November for $90. An update for Professional 9.1 will cost $60, as will a student edition.

Temporal Dynamics Creates High-Speed Database for Linux

While relational databases are a great invention and have brought sophisticated data processing to even the smallest of companies, they are not good at all jobs. They have limits to scalability and are not designed to support truly real-time transaction, sensory, or process control environments. That's why very smart computer scientists created in-memory database management systems, which can scale to petabytes and deliver much more peppy performance.

Temporal Dynamics, a developer of high-performance databases in Indianapolis, Indiana, that was founded a year ago, has announced its entry into this exotic portion of the database market, called Ancelus. This database, which is now available for Solaris and Linux in Ancelus Version 1.3, will be available for HP-UX and Windows in future releases and, presumably, for AIX at some time further down the road.

Ancelus 1.3 can manage datasets as large as 32 terabytes (this is a lot of main memory) and can be extended further to address 8 exabytes (that's 8 million terabytes, and, by the way, no one makes main memory that is large enough to support this, and will not for the foreseeable future). The database has native support for threaded, sequential, and time-based data sets, and allows dynamic administration of the database without having to take it offline, and also sports an online continuous incremental backup feature. You can't turn off the database behind the phone switching system or a nuclear reactor to do a backup, after all.

Avnet, HP Partner to Push Servers in Mexico

The Mexican units of server maker Hewlett-Packard and IT distributor Avnet announced that they have extended Avnet's HP server reseller agreement in the United States to cover Mexico. Avnet has operated a reseller unit in Mexico since 1985, which now has 55 employees; HP says it is the dominant supplier of X86 and Unix servers in the Mexican market, and is looking to Avnet to help it expand that operation. The Mexican reseller agreement between Avnet and HP also covers storage and systems management software. Avnet will be providing financing and tech support services under the deal as well.

Offshore Outsourcing Growing At 20% Per Year, META Says

The offshore outsourcing market is currently a $10 billion business and will grow at a rate of 20 percent per year through 2008, according to META Group, which held an outsourcing conference last week in San Francisco. The growth figure, which means it will reach about $21 billion by 2008, backs up another recent accounting of the growth of offshore outsourcing done by the Government Accountability Office, which found that, from 1997 to 2002, the offshoring of business, professional, and technical services (of which IT is only a part) grew from a $21.2 billion business to $37.5 billion, a 77 percent increase. Eventually, META says, the average enterprise will use offshore resources for 60 percent of its work related to developing and maintaining applications. Application development and maintenance currently accounts for about 30 percent of the outsourcing work sent offshore, which saves organizations about 30 percent, the group says. But the potential savings are much greater. For other types of IT work, which META did not specify, offshore labor can be anywhere from three to five times cheaper, the group says. The current political and public backlash against offshore outsourcing is just one more "item to be managed in the process of moving offshore," META says.