evaluation

Qiao XIN (qxin@cs.ucsd.edu)
Mon, 8 May 2000 22:31:00 -0700 (PDT)

Qiao Xin

Evaluation of the Paper: Why Aren't Operating Systems Getting
Faster

This paper evaluates several hardware platform and operating systems
using a set of benchmarks. Eight benchmarks were run on various
hardware-OS combinations which stress on kernel entry-exit, context
switching, block copy, file systems. The benchmarks suggest operating
systems performance is not improving at the same rate as the base
speed of the underlying hardware. Four potential problem areas are
pointed out.

1) Memory bandwidth is not keeping up with the CPU speed.
2) Context switching both for kernel calls and for process switches is
about 2x more more expensive in RISC than in CISC machines, and the
author has no sure answer to it.
3) Decouple file system performance from disk performance in the
operating system development.
4) In the area of network protocols, the assumptions inherent in NFS
represent a fundamental performance limitation.

>From this paper, we can conclude that the situation can be improved
through improving memory bandwidth and reduce operating systems'
tendency to wait for disk operations to complete.


Evaluation of the Paper: The Interaction of Architecture and
Operating System Design

This paper examines directions in computer architecture and operating
systems, and the implications of changes in each domain for the
other. In this paper, requirements of the operation systems are
discussed related with commercial RISC architectures. It is revealed
that architectures have paid less attention to operating system
requirements, and new operation system design often have overlooked
modern architectural trends. Thus the problem is operation system
performance is well below application code performance on contemporary
RISCs.

Compared with Ousterhout's paper, this paper focuses on the
relationship between essential operating system primitives and
hardware support. Three component of operating system design are
investigated. First, the evolution of operating systems from
monolithic, centralized kernels to a more decentralized structure
requires good communication performance which relies on primitive
functions that do not fully benefit from modern architecture
innovations. Secondly, operating systems requires more memory
management, while new architectures add significant complexity and
latency to the memory management task. Third, thread management seems
to be the area in which architectures and operating systems have moved
at cross purpose.

I think the useful conclusion of the paper is that architects should
pay more attention to operating systems, and operating system designers
should pay more attention to architecture. But answers to some
problems are not given in the paper.