Diagnosis and Performance Monitoring - Deploying and Managing a Cloud Infrastructure - page 129

And because of the growing chasm between processing performance and disk access
times, there is an obvious performance bottleneck. This growing disparity between process-
ing and mechanical performance coupled with the increase of capacity of main storage actu-
ally means only one thing: the high-performance file systems such as those used in cloud
computing systems should employ special caching algorithms or techniques in order to cover
the performance gap created by disk access latencies.
Synchronous Metadata Update
Metadata is still basically a file in of itself, but it's handled a bit differently than other types
of files. It is important to the file system, so it sometimes results in a bottleneck to the sys-
tem, and that is why there is still active research going on dedicated to enhancing metadata
performance. This performance is traditionally slow because metadata has to be written
synchronously. This is called synchronous metadata update, and it happens because the file
system resides partly on volatile buffers and persistent disks where the file system writes
everything to disk as soon as a write command is received. File systems do not update every-
thing synchronously because synchronous updates result in a slow file system, but metadata
is still written in this way.
Metadata is written this way because many believed that file systems that employ syn-
chronous metadata updates (like BSD and FFS/UFS ) are safer compared to file systems that
do not (like the Linux ext2fs ) write metadata synchronously. Let's put that into perspective
with an example.
You are working on a file in the UFS file system for 15 minutes and just recently executed
a save buffer when the machine suddenly crashed. After bootup and the subsequent file sys-
tem check, everything was gone; the autosave file, the backup file, everything either did not
exist or was empty.
This was caused by synchronous metadata update. When the save buffer was executed,
the file was written to the file system; the metadata got written to disk, and the data that
was kept in memory was lost when the power was removed and the subsequent file system
check made this file empty. The autosave file was then deleted even though the data had
already been written on disk because it had no associated metadata. There was also no
backup file written because it was a newly created file and the first save.
Soft Updates
A soft update is the alternative to sequence-dependent or synchronous updates and
write-ahead logging. Write-ahead logging writes all modifications to a log before they
are applied to the actual storage (on disk, such as in a database). Write-ahead caching
ensures atomicity of operations, which essentially means that an update is either com-
pletely written or completely discarded (in case of a failure), thus ensuring system con-
sistency. This method allows the safe use of write-back caching; that is, modifications
to data in the cache are not copied to the cache until absolutely necessary. This enforces
and tracks metadata update dependencies to ensure that the disk image remains in a
consistent state. This method improves metadata performance by combining multiple
updates into fewer background disk writes.