MIT Randomizes Tasks To Speed Massive Multicore Processors

itwbennett writes Researchers at the Massachusetts Institute of Technology have created a data structure that they claim can help large multicore processors churn through their workloads more effectively. Their trick? Do away with the traditional first-come, first-served work queue and assign tasks more randomly. The SprayList algorithm allows processors with many cores to spread out their work so they don't stumble over one another, creating bottlenecks that hamper performance.

Comments

Re:I don't know enough about this stuff

[mwvdlee]

Results don't come in out-of-order. Imagine two variables, A and B, each undergoes a number of calculation steps which don't refer to the other variable. I.e. A=A+5/2*13-29 and B=B*B*3+12/N, then finally adding them together as Z=A+B. Normal execution would first do all the calculations for A then all the calculations for B, then finally Z. Out-of-order execution would calculate both A and B simultaneously, wait for both to finish, then calculate Z. Out-of-order execution involves a lot of this type of waiting, but since it's waiting for just the slowest calculation instead of the sum of both the slowest and fastest calculation it ends up being done sooner. If things cannot be calculated like this, an out-of-order capable processor will simply do things in-order.

At least that's how I understand it at a very abstract level.