evals for 5/18/2000

Aaras P VASA (avasa@cs.ucsd.edu)
Thu, 18 May 2000 09:52:10 -0700 (PDT)

lightweight remote procedure call
this paper talks about LRPC designed so as to lower the cost of
communication between protection domains on the same machine
a predominant occurence

traditional RPC incur a high cost or sacrifice safety to
achieve better 'same domain' performance
LRPC uses large-grained parallelism of RPC and control transfer
and communication model of capability-based systems
their resesrch shows that cross-domain small parameter transfer
dominates and RPC has high overhead for this.
LRPC uses Arguments-stack to lower copying, privately mapped
execution-stacks for protection and run-time simple stubs few
things which help to reduce cost

i think LRPC is a good option.they are even more suited for
multiprocessor operation.their argument is based on the fact
that communication occurs mainly in the same domain and with
few parameters, though not always true, i agree with.
it doesn't work well with large parameters and has the additional
problem of complicated handling of failures.

Active messages: a mechanism for integrated communication and
this paper tries to use active messages so as to overlap
communication and computation and lower the start-up cost

active messages are an asnchronous communication mechanism which
try to generalise the hardware functionality.each
message contains at its head a message handler to get the messages
out of the network and into the computation .

active messages help in eliminating buffering.the parallel
machines communication try to eliminate the OS intervention
and security is less important as compared to RPC.
hardware support for active messages is needed.this paper
argues that traditional programing models dont overlap
communication and computation too well,which i agree.
but i am not convinced about the hardware
paper argues hardware support for communication is counter-productive
and also states a huge amount of hardware support needed for
active messages