I've recently been stumbling into interesting papers on new ways to wire up datacenters to guarantee properties like high bisection bandwidth without the need for expensive networking hardware like big second level switches. I don't normally read a lot about networking but some of the stuff I've been working on lately has given me a peek into the SIGCOMM, NSDI world ...
CamCube, BCube, and DCell all seem to be exploring ways of giving up the big expensive switch at the top of a tree-structured network in favor of multiple NICs on each server and/or multiple smaller switches connected in interesting topologies. I suspect this is still very early stage stuff, and many problems such as cabling, servicability, availability NIC-to-NIC transfer without burning CPU etc need to be solved before it becomes interesting to a practitioner. But the papers do offer an intriguing read for someone who normally follows the database and systems communities.
In the Hadoop/MapReduce context, this could change how we think about scheduling tasks, scheduling transfers when multiple MapReduce jobs are running. One of the papers actually talks about how Hadoop workloads could be affected by these new topologies. One criticism against these approaches in the context of Hadoop clusters is that they probably only get interesting for mid-to-large clusters. Monash reports that the median Hadoop cluster is about 30 nodes and that the average is about 200 nodes (numbers courtesy Omer Trajman of Cloudera). At these sizes, there are cheaper and easier ways to wire up a Hadoop cluster. Fun papers nevertheless!
No comments:
Post a Comment