221

Jamison Collins (indecent_and_obscene@yahoo.com)
Thu, 1 Jun 2000 03:34:44 -0700 (PDT)

Fast File System for Unix

This paper presents a reimplementation of the Unix
File System. The old
implementation has failed to scale in performance with
faster processors
and disks, and this new implementation seeks chiefly
to regain disk bandwidth,
as well as add often requested features.
The chief performance loss in the old system is due to
overly small file
blocks. Thus, one of the key modifications made by
this reimplementation
is to increase the size of the file blocks. However,
simply increasing the
size will result in a dramatic increase in the amount
of wasted disk space.
As a side note, the author doesn't know how well he
had it back then... I
remember when people were waiting for Win95B to be
released because we
were stuck with a 32k block size because of the
inefficiencies in FAT16.
Anyway, in order to regain this wasted space, the file
system allows block
splitting so that a full block doesn't need to be
allocated to a file
which doesn't need the entire block for data.
Secondly, in order to target disk seeking delays,
files are organized into
cylindar groups in order that the file information for
a group of files
can be located closely together. Finally, several
useful featueres were
added.
These changes are evaluated in terms of the
performance features that were
targetted: does the FS allow a larger block size
without a great increase
in the amount of waste, and is disk bandwidth
improved? To both the answer is
yes.
This paper presents an incremental change to an
existing file system structure.
While they are convincing that each mentioned change
is effective in targetting
the particular problem it is targetting, it is not
totally clear that the issues
targetted are severe problems, or that the degree of
improvement that was
achieved was totally necessary. But I guess I'm just
used to seeing IPC
numbers are the end of the paper. In any case, the
ideas presented are
interesting.

Log structured FS

This paper presents a log structured filesystem.
Similarly to the last paper,
this file system attempts to overcome inefficeincies
due to the original design
of the Unix FS. However, this system makes radical
changes to the FS structure
in order to achieve significantly higher performance.
One key trend that is identified is that disk caches
will allow most read requests
to be satisfied by memory. Thus, write traffic will
become increasingly
important for disk accesses.
Two important problems are identified--the number of
seeks necessary to
create a file and that writes typically occur
synchronously.
The segment concept is introduced to target both of
these. By grouping
writes into segments which are written sequentially
across the entire disk,
creating a file only requires writing information into
a segment and updating
a global map. This sytem, however, requires
maintenence because eventually
data will be stored in every disk position, even
though much of the data
has been logically(but not physically) deleted or
overwritten. To address
this, segment cleaning is discussed, which is a
procedure by which this
old and useless data is removed from existing
segments. A series of results
are presented in order to determine the optimal
cleaning strategy. Finally
crash recovery is discussed and microbenchmark results
are given to demonstrate
that performance has been improved in the key areas
that were targetted.
These modifications are primarily evaluated through
microbenchmarks, which
demonstrated that highly effective nature of the
mentioned filesystem.
Now, this was good research. A reevaluation of an
existing, deeply help design
belief, with the result being that a fundamentally
different structure. This
is the kind of papers I like to read because the
results gained provide a
key insight into the fundamental nature of filesystems
and the bottlenecks
better than a paper which simply modifies existing
structures.

=====
I have often wondered... what if we discovered
that we were threatened... by an outer power,
from outer space... from another planet?
Wouldn't we all of a sudden find that we
didn't have any differences between us at all?
That we were all human beings?

__________________________________________________
Do You Yahoo!?
Send instant messages & get email alerts with Yahoo! Messenger.
http://im.yahoo.com/