Sequencing the SARS Virus
By the time we bought our first set of IBM x330 servers, now part of a 168-CPU cluster (Figure 7), the 1U platform was on the verge of entering the commercial off-the-shelf (COTS) category and starting to enjoy COTS prices. Beige boxes are no longer used for distributed computing. Heavily loaded production subsystems, like Apache and MySQL, are housed on IBM's 4U x440s, which are eight-way hyperthreading Xeon nodes with 8GB of RAM. These boxes are running SuSE 8.1—one of the few distributions that supports IBM's Summit chipset. The x440 is a NUMA machine with 32MB of L4 cache per four-CPU module, and without IBM's Summit patches only presents two CPUs to the kernel. SuSE's 2.4.19 derived kernel with bigmem+Summit support makes it possible to use all eight CPUs and 8GB of memory. Even without the advanced NUMA scheduler code now in the 2.5 series kernels, these x440s have been real workhorses allowing us to run eight BLAST processes concurrently with enough RAM to cache the entire human genome in shared memory. Anyone who claims Linux isn't ready for Big Iron is in for a surprise.
As we rapidly grew, the NFS subsystem was beginning to show problems. In particular, machines were crashing with some combinations of NFS server/client versions. Although in our experience NFS clients are robust, room for improvement exists with current Linux NFS services. Our fastest NFS server, an IBM x342 (2xP3-1.24, 2GB RAM) cannot handle more than 4,000–6,000 NFS ops/second, especially during a large number of parallel read/writes from our cluster. To address the performance limits, we acquired a NetApp FAS960 Filer (Figure 7). With 10TB of raw storage (5x14x144GB), the filer has reached 30,000 NFS ops/second. Despite NFS issues, our original VAR900 production file server (Figure 2) was the poster child of stability and reached an uptime of 394 days in February 2002 before it had to be rebooted for upgrades.
The first set of Tor2/SARS sequence data was available for our informatics group to analyze on Friday evening, April 11, 2003. To verify our sequence reactions, we checked it for contamination. A BLAST search allowed us to determine the closest match in the public proteomic and genomic databases. To our relief, the best match was to bovine coronavirus (Figure 8), indicating that we were sequencing something related to coronaviruses. The sequences of these viruses end in a string of As, and when we saw sequence reads ending in a poly-A tail we were confident that this was one end of the genome.
The x330s and an x440 were used to analyze and assemble the SARS data. The genome is not very large, and the assembly took less than 15 minutes on a single CPU. In comparison, the first public assembly of the human genome, 300,000 times the size of Tor2/SARS, was done at UCSC and took four days on a 100-CPU Linux cluster.
By Saturday, April 12, 2003, at 2:25AM, we completed our seventh build of Tor2/SARS, and this assembly was frozen as the first draft. It was imported into AceDB to visualize alignments to other known protein sets for validation (Figure 9). We spent Saturday validating the assembly, which was posted later that day to our x440 public Web server using a custom CMS system running under Zope/Plone.
The sequence of Tor2/SARS has identified a fourth novel group of coronaviruses and provides the necessary information to develop diagnostic tests and, possibly, therapies including a vaccine. Linux has made it possible to get our work done without spending a fortune on hardware or software. Using commodity hardware has minimized depreciation loss due to long implementation times. We'll be watching for new bugs to catch, and in the meanwhile, our MySQL database is open for sequencing.
- New Products
- New Products
- Integrating Trac, Jenkins and Cobbler—Customizing Linux Operating Systems for Organizational Needs
- Tech Tip: Really Simple HTTP Server with Python
- Dialog: An Introductory Tutorial
- RSS Feeds
- Non-Linux FOSS: Remember Burning ISOs?
- Returning Values from Bash Functions
- EdgeRouter Lite