Conclusions

This paper focused on the kernel issues involved in building Direct Access File System servers. A range of issues was addressed drawing from our experience in building such a kernel server for FreeBSD. We described the current server structure using existing interfaces with minor kernel modifications. Performance results show that our current DAFS server prototype implementation can offer high performance file service for memory workloads over a 1.25 Gb/s network.

A problem with existing blocking kernel interfaces is that DAFS and other kernel servers using them experience the overhead of having to associate a process context with each I/O request. This overhead is expected to become more pronounced in multi-gigabit networks. Our proposed additions to the vnode interface offer support for asynchronous file I/O with integrated network and disk event notification and delivery.

We presented design possibilities for integrating a programmable RDMA-capable NIC with the FreeBSD VM system. This support will allow a DAFS server to export its entire VM cache over the network in the face of client-initiated RDMA operations and server paging activity. We are currently planning an implementation that embodies such a design and can be used to support Optimistic DAFS.

Finally, we have found that the existing BSD network driver model is inadequate to support the needs of memory-to-memory NIC devices and a new model is needed.