e2fsck IO parallelization
I implemented parallel readahead of metadata in e2fsck using a pretty
hysterical combination of fadvise(), read(), and blindfolded buffer
cache manipulation. It gets about 50% improvement in elapsed time on
a RAID5. Current e2fsck readahead
My friend Akkana
followed my advice to use the noatime mount option to cut down on
writes to her file system, but then mutt couldn't tell which mailboxes
had been read since they were last modified. So I wrote the relative
lazy atime patch, which only updates the atime if the old atime was
older than the ctime or mtime. It turns out this is terribly useful
for distributed/clustered file systems as well, since metadata writes
are generally very expensive and the alternative is either no atime,
or some kind of annoying tunable atime quantum. Currently in -mm,
rewritten by Andrew Morton to be slightly less turgidly dense. Now
with accompanying patch for the mount utility (although you can still
just use the test program to test it if you don't want to download and
Relative atime kernel patch against 2.6.18-rc4
Relative atime userspace patch against util-linux 2.13-pre7
Test script to test relatime
Bare bones mount program to remount with either relatime or lazyatime (depending on kernel patch)
Older lazy atime patch, which updates the atime in memory but doesn't bother writing it to disk. The major problem with this is that atime can go backwards if an inode is evicted from memory.
Lazy atime patch against 2.6.18-rc3
The default malloc() threshold for allocating using mmap() versus
brk() is designed for 32-bit systems with 2001 memory capacity, and
winds up doing too many allocations using mmap(). This is especially
harmful when the application is doing an alloc/write/free cycle, as
the memory has to be cleared unnecessarily in between each free and
alloc. Arjan van de Ven and I wrote a patch to glibc that dynamically
adjusts the threshold based on the size of free()'s.
Update: Merged as of August 24, 2006 (with some additional tweaks).
Dynamic mmap threshold
ebizzy workload generator
I needed a workload that represented a class of applications fairly
common in web application servers (think search and indexing
services). Basically, a lot of threads are all doing the same kind of
thing to a big (many GB's) chunk of memory: pick some chunk of memory,
make a copy, look up some record in it using binary search, free that
memory again. Microbenchmarks don't cut it, because the performance
issues only kick in when everything is combined. ebizzy tries to
replicate this for system tuning purposes. Licensed under the GPL.
The source is now hosted on Sourceforge.
OLD 0.1 release
Ext2 fsck reduction project
These patches are intended to reduce the average time spent on fscking
ext2 file systems. For more info, see the paper (draft).
Latest patch for fswide dirty bit
Port of ext3 reservations to ext2 Martin Bligh and Andrew Morton cleaned this up and it's floating around in one of Andrew's trees somewhere; ask for it.