Paper review

Bryan Wang (
Wed, 31 May 2000 23:58:49 -0700

Bryan Wang, 5/31/00

The Design and Implementation of a Log-Structured File System

Rosenblum and Ousterhout are trying to improve the performance of disk
i/o through the use of log-based file systems. They recognized that
disk access times must improve with CPU speeds or else performance
will be limited by disk drives. Also, disk i/o times are dominated by
writes, because most reads are taken from large main memories. A
log-based file-system will allow a large chunk of write data to be
buffered and written all at once, effectively eliminating disk seek
times required by random-access writes.

Segment cleaning was the biggest hurdle they faced. They performed
several simulations to determine optimum segment cleaning policies,
and ultimately decided on a cost-benefit algorithm. Compared to
UNIX filesystems, the log-structured FS is much faster for working
with small files, and also allows the system to recover more quickly
from crashes.

Never having heard of a log-based file system, this is all new to me.
I found the paper to be very well-written and easy to follow.
It was probably the easiest paper to comprehend out of any we've read
in class. And I found it hard to argue with the merits of the
log-based file system, so I enjoyed the paper.

Bryan Wang, 5/31/00

A Fast File System for UNIX

McKusick et al. wished to improve the bandwidth of the "old" UNIX file
system without changing the existing abstractions (to maintain
software compatibility). They used a system of large blocks (4096
bytes instead of 1024 bytes) and block fragments (512 bytes minimum,
or the sector size) to create a more efficient file system.

A very interesting yet complicated technique they chose to include is
file system parameterization. The system is analyzed (CPU speed, i/o
interface, and disk rotation and seek times) to create an omtimum
layout for each particular system configuration. While this is
certainly a worthwhile optimization to consider, intuitively I prefer
a solution that depends little on underlying hardware.

The team certainly suceeded in their goal to create a more efficient
file system than the old system. But after reading both this and the
LFS paper, I prefer the LFS system for both its lower complexity and
superior performance.