| Cluster 2005 | START ConferenceManager |
In this paper, we present a new, analytic model that captures the benefits of protocol offload in the context of high performance computing systems. In contrast to the LAWS model, our model emphasizes communication in terms of messages rather than flows. This allows us to consider benefits associated with the reduction in message latency along with benefits associated with reduction in overhead and improvements to throughput.
We show how our model can be mapped to the LAWS model and we show how it can be used to contrast full offload (TCP offload engines) with other approaches, e.g., interrupt coalescing and protocol bypass. Full offload moves all of the protocol processing onto to a NIC, reducing overhead and latency while providing full throughput. Interrupt coalescing reduces overhead by reducing the number of interrupts due to communication. Full TCP offload engines require substantial NIC resources and therfore drives up the cost of a NIC. In contrast, interrupt coalescing minimizes the need for additional NIC resources, but increases jitter in message latency. In protocol bypass, protocol headers are queued for later processing while the data is transmitted directly to the application. This requires only a modest increase in NIC resources. Using our model, we demonstrate that protocol bypass retains most of the benefits associated with full offload and, as such, represents an important middle ground between interrupt coalescing and full offload.
| START Conference Manager (V2.49.7) |
| Maintainer: rrgerber@softconf.com |