evaluation

Qiao XIN (qxin@cs.ucsd.edu)
Wed, 17 May 2000 23:55:44 -0700 (PDT)

Qiao Xin

Evaluation of the Paper: LRPC

This paper describes Lightweighted Remote Procedure Call which is a
communication facility designed and optimized for communication
between protection domains on the same machine. The reason for the
appearance of LRPC is in conventional RPC system local communication
is treated as an instance of remote communication and simple
operations are treated as the same class as complex ones, therefore
incurs unnecessarily high costs.

LRPC combines the control transfer and communication model of
capability systems with the programming semantics and large-grained
protection model of RPC. By reducing overhead of same-machine
communication, it achieves performance as well as retains safety and
transparency of conventional RPC. I think some techniques are worth
mentioned here.

In LRPC, the client's thread executes the requested procedure in the
server's domain; Shared argument stack between client and server
eliminate redundant data copying; simple and stylized LRPC stubs which
consists of mainly move and trap instructions results in good
performance improvement; LRPC avoid shared data structure bottleneck
and also reduces context-switch overhead by
caching domains on idle processors in the multiprocessor case. By
exploiting simple control transfer, simple data transfer. simple stubs
and multiprocessor, LRPC represents a viable communication alternative
for small-kernel operating systems.

Evaluation of the Paper: Active Messages

This paper discusses a communication design for large-scale
multiprocessors. It addresses the problem in existing message passing
mechanism and message driven machines. Existing message passing
multiprocessors have high communication costs, while message driven
machines demonstrate poor processor cost/performance.

Active messages is an asynchronous communication mechanism intended to
expose the full hardware flexibly and performance of modern
interconnection networks. The basic idea is to overlap
communication and computation and to reduce communication
overhead. Under the active message model each node has an ongoing
computation which is punctuated by asynchronous message arrival. A
message handler is used to extract the message data and integrate it
to the computation. It has the efficiency of eliminating buffering
beyond network transport requirements, simple scheduling of
non-suspensive message handler and arbitrarily overlap of computation
and communication.

Other benefits of active message model includes it utilizes the full
capability of the hardware, captures the essential functionality of
message driven machines with simpler hardware mechanisms and is
independent of any programming model.

Questions about the paper is how is active message related with RPC.