Faster Cores, but a Slower Network

Arthur Cole

Did you ever have one of those dreams where, no matter how fast you run, you don't seem to be getting anywhere?


That's the feeling a lot of enterprise managers are likely to experience in the coming age of dual cores, quad cores, multicores and other enhancements to CPU performance. As the folks at HPCwire pointed out in an editorial last week, there is a growing gap between theoretical and realized performance as each boost in processing power adds a new layer of complexity to the overall network.


One of the immediate problems with new multicore servers is that while lightning-speed processing may take place within the server, there will likely be a bottleneck on the interconnect. To that end, a company called Mellanox has developed an Infiniband I/O that operates at up to 20 Gbps. The ConnectX Host Channel Adapter delivers an application latency of only 1 microsecond.


Another way around the I/O squeeze is to bypass it altogether. That seems to be the approach that NEC and Stratus haven taken with a new pair of fault-tolerant servers. While the two devices maintain their own OSs and applications, they share a common clustering and replication stack on a common memory level, avoiding the need to transfer data to and fro should one unit fail. From a network perspective, they are a single device running mirrored copies of the same software. This makes service interruptions a thing of the past because it's as if a failure has never taken place.


But even while networking technology plays catch-up with CPU designs, there's always another advancement waiting in the wings. While it may be a long way from commercial applications, it might be worth taking a look at the polymorphic computer that Raytheon recently developed for the Defense Department. Essentially, it's a device that can reconfigure itself on the fly to jump from, say, a streaming workload one moment to parallel the next. The company says its array of 6 processors provides 64 gigaflops with more than 60 GBps of memory bandwidth.


Let's see them build an I/O that can handle that.

Add Comment      Leave a comment on this blog post

Post a comment





(Maximum characters: 1200). You have 1200 characters left.




Subscribe to our Newsletters

Sign up now and get the best business technology insights direct to your inbox.