Paper Evaluaiotns

Kiran Tati (ktati@csag.ucsd.edu)
Tue, 16 May 2000 08:13:34 -0700

Performing Remote Operations Efficiently on a Local Computer Network.
The main goal this paper is to design instructions to access the data
on remote machines, which provide a fine grain control for accessing
data on remote machine. I think the idea is simple extension to tightly
coupled multiprocessors with a shared memory into the local area
network. The paper claims this design provide more flexibility to the
application programmers to implement the remote operations because these
small remote instructions are remote procedure calls in the assembly
language. One can easily design the higher-level language to support
more flexible to the application programmer for accessing the remote
machine services. Since data transfer in one operation is very small so
that one can easily optimize the overhead associated with the network
data transfer. Another advantage is that since the data transfer is very
low it can be implemented at a device controller level so that one can
reduce the number of context switches. The paper also presents how
these small remote data access instruction can be designed to get the
various levels of reliability and synchronization.

The paper just compares the ordinary network protocols with for the
small data transfers but these networks are designed for the large data
transfer in mind (at least not for a single word data transfers). These
general-purpose network protocols are designed based on the assumption
that the network links varying in the speed and reliability (may be very
few hundred bytes per seconds to megabit per second, and not only the
local area networks). I think the best comparison might be some tightly
coupled shared memory access times. Of course these systems doesnít
scale well because the number of instructions needed to do a small job
requires a lot of such small instructions and in this case the one can
replace the all such small requests with very big request.

Implementing Remote Procedure Calls.
The main goal of this paper is to ease the distributed systems
programming easy by implementing the remote procedure call mechanism,
which is simple extension to the simple procedure call. The other goals
of this paper are the efficiency of this RPC as well as the security.
The paper also explains the design decisions they took when they
implemented the RPC, which the previous papers didnít discuss
(communication failures, binding, security and integration with the
programming languages). The binding of the remote interfaces is
exposed though the Grapevine distributed data base. Usually the name of
the interface consists of type as well as some information about the
instance of that type. The server of the interface is assigned a unique
32-bit value that a client receives when he resolves the interface name
and this value is used to know about the server failures. Using the
Grape vine database also provide some form of security. Another
important decision is that use of specialized transport protocol for the
RPC mechanism to reduce the connection establishment overhead and state
information maintained by both sides of the connection. Piggy backing
are used to reduce the number of packets transferred. The transport
protocol is heavily optimized for the small requests. The server process
on the remote machine is not terminated when all the requests are over.
The network device handler can wake up the waiting process to
immediately respond the requests. The special link layer protocol is
used to reduce the overhead associated with the layered communication
protocols.

--
Kiran Tati (Graduate Student)            Kiran Tati
UCSD/CSE AP&M 4402                       9450 Gilman Drive
9500, Gilman Drive, Dept. 0114           UCSD#922563
La Jolla, CA, 92093-0114                 La Jolla, CA 92092-2563
Phone:  858 5345486 (Office)             858 552 9291 (Home)
E-Mail: ktati@csag.ucsd.edu

Office Address: AP & M 6426