evals for 5/4

Aaras P VASA (avasa@cs.ucsd.edu)
Thu, 4 May 2000 09:05:08 -0700 (PDT)

the distributed V kernel

the goal of this paper is to prove that a distributed sytem with diskless
workstations and V message facilty perform nearly as well as a more
specialized file access protocol

The V kernel consists of many small processes communiating by synchronous
fixed-size request-response messages.it ditinguishes between small
messages and data transfer facility.

the paper emphasizes the fact that the cost of file access using V kernal
is nearly the same as other means.one important thing to note in the
comparisons is that all the acceses are remote accesses . while in the
case
of disks there would have been local accesses also.
they assume the fact that the network is never loaded , not a practical
assumption at all.
file server processing will be a problem according to me.
V doesn't implement streaming which can be really beneficial in some
cases.
overall although general interprocess communication might perform well
but efinitely not as well as suggested in the paper

The sprite Network Operating System
sprite is a network operating system developed because of the changes
in networks, physical memories and multiprocessors.
the goal was to provide mechanisms to capitalize on these 3 trends.

The sprite kernel uses specialised RPC for communication between kernels
it provides a tranparent network file system which is implemented as a
collection of domains, but appears as a single hierarchy.the file system
uses prefix tables.it caches file data on both server and client
nmachines.
sprite also provides process migration.
sprite thus allows the sharing of disk storage , physical memory and
processing
power

i think this is a well written paper. the idea of the kernel being
multithreaded seems a good one although it increases complexity.
Prefix tables also seem a good idea to mange files.
the main thing which was not entirely convincing to me was the drawbacks
of the cache consistency mechanism. i think this particular mechanism
does limit the performance. even in the performance measurements the
overhead
of cache consistency is ignored.
also protection is a problem as kernels are expected to trust each other