cse221: paper evals

Octavian Luca (oluca@cs.ucsd.edu)
Thu, 18 May 2000 05:28:39 -0700 (PDT)

Lightweight RPC

This paper discusses advantages of using lightweight RPCs that are
optimized to take advantage of the fact that most communication occurs
between protection domains on the same machine. This paper differentiates
itself from previous work in its focus on optimization of communications
between protected domains on the same machine. This work describes a
significant way to improve common operations resulting in significant
performance improvements.

The paper outlines other methods for acheieving fine-grained protection
such as capability systems and their corresponding disadvantages such as
thir inherently difficult programming interfaces. Four reasons are given
for the increase in performance of LRPC over other techniques: simple
control transfer, simple data transfer, simple stubs and design for
concurrency. The authors argue the merit of their new method successfully
by explaining the advantage of their method over previous ones and
exemplifying their arguments with measurements. They also argue that the
uncommon cases are also handled well by LRPC, proving that the new method
yields satisfactory results under all circumstances.

I thought the paper was well written and i liked the way they explained
their idea and then proved that it is advantageous to previous ones. This
paper seems to imply that further work is likely to continue in this area
since this research has yielded significant performance improvements.

Active Messages

This paper introduces the mechanism called active messages and discusses
how it can be used to implement dynamically scheduled languages for which
message driven parallel machines were designed. The research reflected by
this paper differs from previous work in its focus on a low level
mechanism meant to achieve high processor efficiency by balancing
communication and compute times. The importance of this topic is derived
from the perpetual need to improve the efficiency of parallel
architectures and improvement of processor cost/performance ratio.

The paper first examines message passing architectures based on current
ideas and shows that send/receive programming models poorly utilize the
hardware resources. Then the researchers look at message drived designs
and show that expanding the power of message driven processes beyond
active messages is costly and not necessary to implement the desired set
of parallel programming languages. Lastly, they discuss some of the
issues relating to better support for active messages in hardware. Active
Messages differ from RPC in that the role of the active message handler is
not to compute data, but rather do extract the data from the network and
integrate it with ongoing computations with minimal cost. The key
optimization of active messages when compared to send/receive is the
elimination of buffering and the overhead associated with it.

The authors claim that using active messages to guide software design, it
is possible to improve current message passing machines in a so-called
"evolutionary fashion." According to the paper their work is similar to
contemporary defelopments of RPC mechanisms as both try to reduce
communication layer functionality and to optimize for the common case.
The main conclusion I gathered from this paper is that the fundamental
issue in designing a balanced machine are providing the ability to overlap
communication and computation and to reduce communication overhead. The
authors have argued that active messages minimize software overhead in msg
passing machnes and use the full potential of the hardware.