CSE 221: Paper Evaluations

Renata C TEIXEIRA (teixeira@cs.ucsd.edu)
Mon, 8 May 2000 22:26:21 -0700 (PDT)

Title: "Why Aren't Operating Systems Getting Faster As Fast As Hardware?"

This paper evaluates several hardware and software configurations and
identify why the operating systems speed is not scaling at the same
rate as basic hardware speed. The evaluations were performed using a
set of "micro-benchmarks", which measured specific features of the
hardware or the operating system, and one "macro-benchmark", which
gives a better idea of overall system performance.

The benchmarks indicated that even though the RISC machines were
generally faster than the CISC machines, they were not as fast as
their MIPS rating would suggest. Memory-intensive benchmarks show
that this kind of applications are not likely to scale well in
RISC systems. Benchmarks comparing the file systems show that
Sprite is faster than other operating systems, particularly, in
the remote accesses and that Unix and its derivates are more
disk-bound, some operations cannot be completed until I/O is
completed.

Based on the results, the paper suggests that memory bandwidth should
increase, that it is important to decouple file system performance from
system performance, and that NFS assumptions represent a performance
limitation.

The paper presents an interesting analysis, showing that the
interaction between the operating system and the hardware
architecture is very important to the overall performance
of the system. The performance evaluation, however, was only
based on benchmarks, which not necessarily represent real usage
of the system, and it doesn't quantify the relative usage of
the functions analyzed and the other functions that benefit
from the processor performance only.
------------------------------------------------------------------------
Title: "The Interaction of Architecture and Operating System Design"

This paper examines the influence of new hardware technologies
on operating systems and vice-versa. The paper compares the performance
of several modern microprocessors executing a set of primitive
operating system functions. The evaluations were performed considering
three major components of operating system design: interprocess
communication, virtual memory and threads.

Interprocess communication performance relies on primitive functions
like: system calls, interrupt handling, and data coping. The paper
shows that these functions do not benefit from RISC architecture
either because they are memory-intensive or because they are no
longer performed in hardware (like in CISC architectures) but
in software and they have to manage new features.

Operating systems are making more demands of the virtual memory
system for functions like copy-on-write and distributed virtual
memory. New architectures virtual memory system, however, are
more complex and hence more difficult to manage.

Larger register sets on newer processors make context switch
of threads inefficient. In addition, new architectures assume
that procedure calls are much more frequent than context switch.

I think that the paper presents a detailed evaluation of how
operating systems perform on new hardware architectures.
It describes specific features of the hardware that are
limiting the improvement of performance on RISC machines. It also
identifies the tendency of new operating systems to be decomposed
and shows that in these new systems the number of the
primitive functions considered in the paper will increase.

Renata Teixeira