HPI-DC'05 Panel:
                 The Future (of) RDMA

RDMA technology is becoming more wide spread and available in 
network interconnects.  The low latencies achievable through 
RDMA are very desirable.  For adoption in high-end parallel 
computing, issues related to buffer management, scalability, 
and standards need to be considered.

From each panelist we seek their opinions and ideas on the 
following questions:

1. Buffer management
Receiver-side buffer management has shown to be successful 
and appropriate for high-performance computing. I.e. the 
receiving application specifies where it wants incoming data 
to land. At the same time zero-copy, OS-bypass, and independent 
progress of the application are important too.

How does RDMA fit into this world? Specifically, how can MPI get 
all the benefits offered by RDMA and achieve zero-copy, OS-bypass, 
and independent progress? Is this possible now, in light of receives 
that may not have been posted yet, or applications doing long 
computations without calling the MPI library? Would changes in 
RDMA technology or APIs make this integration easier and more 
efficient?

2. Scalability
With systems of 10,000 nodes and more, scalability becomes an 
important issue. The number and size of buffers needed on a node 
for RDMA should not grow linearly with the number of nodes in 
the system. Similarly, the amount of state, for RDMA keys for 
example, should also not grow linearly with the number of nodes 
in the system.

RDMA technologies and APIs designed for clusters of a few 
hundred nodes may not scale to tens and hundreds of thousands 
of nodes. What can be done to make RDMA a viable player in 
networks for these high-end machines?

3. Future of RDMA (standards)
Where is RDMA going and how can it benefit HPC?  Will there be 
standards that can be easily adapted by traditional programming 
models; e.g. MPI? Will the mass-market dictate standards and 
technology choices that are detrimental to HPC?