|M.Sc Student||Talmor Yoram|
|Subject||Higher Performance DSM by Peaking Network Utilization|
|Department||Department of Computer Science||Supervisor||Professor Assaf Schuster|
In this thesis, we describe a new architecture we developed to provide better, more highly integrated network support for
Millipage, our software DSM (Distributed Shared Memory) system. Our novel Multiview technique enables an efficient implementation of page-based fine-grain DSM with no false sharing. Our original Multiview-based DSM system used Fast Messages (FM) as its communication library. However, performance measurements proved that FM is not well suited for our unique requirements.
We took a two-phase approach to improve our system performance. In the first phase, we changed FM to be signal enabled instead of polling. This means that a computation DSM node allocates resources for message retrieval only when a message is ready to be processed and no resources are allocated to polling the network for the incoming message.
This enhancement improved the performance of Millipage in several benchmarks. Still, other benchmarks showed that we had additional bottlenecks. In addition, experiments showed that network utilization was low, because we use small messages (we send only the information required by other nodes, not the full system page). This led us to our second phase, during which we migrated parts of our DSM layer to be executed in the network layer context, thus off-loading the node CPU.
This approach improved the performance of Millipage significantly, and it enabled us to achieve high speedups with a wider set of applications. Although this design approach is hard to implement, the reward is high: higher performance DSM, with new research opportunities and better speedups throughout our benchmark suite.