diff options
Diffstat (limited to 'community')
| -rw-r--r-- | community/gsoc/project_ideas/object_lookups.mdwn | 42 |
1 files changed, 42 insertions, 0 deletions
diff --git a/community/gsoc/project_ideas/object_lookups.mdwn b/community/gsoc/project_ideas/object_lookups.mdwn new file mode 100644 index 00000000..462f6a12 --- /dev/null +++ b/community/gsoc/project_ideas/object_lookups.mdwn @@ -0,0 +1,42 @@ +[[!meta copyright="Copyright © 2013 Free Software Foundation, Inc."]] + +[[!meta license="""[[!toggle id="license" text="GFDL 1.2+"]][[!toggleable +id="license" text="Permission is granted to copy, distribute and/or modify this +document under the terms of the GNU Free Documentation License, Version 1.2 or +any later version published by the Free Software Foundation; with no Invariant +Sections, no Front-Cover Texts, and no Back-Cover Texts. A copy of the license +is included in the section entitled [[GNU Free Documentation +License|/fdl]]."]]"""]] + +[[!meta title="Improved system object lookups"]] + +The Hurd currently uses its ihash library ([[hurd/libihash]]) as a generic +container for various objects. While it does its job, it has been reported +to suffer from high collision rates. In addition, the "one size fits all" +approach contributes to slow things down. One particular use case is looking +up an object from a Mach port name, which basically translates to getting the +file or socket associated with a file descriptor in traditional Unix systems. +It's particular because there are actually two lookups for each object, the +first being finding the Mach port from a client port name, which is done in +the GNU Mach kernel, and the second being finding the server object from a +server port name. The best strategy would probably be to directly associate +the address of an object to the receive right of its port, eliminating the +need to look up again, but this is quite an intrusive change in the code base. +For the time being, optimizing lookups would already be an improvement. + +The goal of this project is to increase system performance by speeding up +object lookups, with a particular focus on name-to-object lookups. Note that +there is little room for improvement in the kernel name-to-port lookups because +of the various optimizations IPC has received in the past. Looking up server +objects from port names could use an algorithm highly tuned for this task, +perhaps with better locking (shared/exclusive instead of always mutually +exclusive for example). Then, the libihash algorithm could be replaced with a +better one, not necessarily a hash based one, to improve all the other users. + +This task requires proper knowledge of data structure algorithms, taking into +account machine properties such as processor caches, as well as the appropriate +skills in C and assembly to check the generated code. Being able to perform +accurate measurements in a system that lacks modern profiling tools would also +be helpful. + +Possible mentors: Richard Braun |