Home » Mailing list archives » dragonflybsd-kernel » 2008 » June

dragonflybsd-kernel mailing list

FromSubjectsort iconDate
Dan M
Re: HAMMER Update 13-June-2008 - HEADS UP: MEDIA CHANGED AGAIN
Neat. Thanks for the lengthy write-up. -- Dan
[ message continues ]
Jun 16, 4:37 pm 2008
Matthew Dillon
Re: HAMMER Update 13-June-2008 - HEADS UP: MEDIA CHANGED AGAIN
I can never keep all the variations straight, I need a Wiki reference every time to get the name right :-). These are all minor variations of B+Trees. The B* variation keeps nodes 2/3rds full instead of half full. It is an optimization which is intended to maintain higher fill ratios to make better use of system caches. A Dancing tree ... hehehehe. Looks like Hans invented that one. I'm not sure it even counts as a variation of a B+Tree. It doesn't ...
[ message continues ]
Jun 16, 2:37 pm 2008
Matthew Dillon
Re: HAMMER Update 13-June-2008 - HEADS UP: MEDIA CHANGED AGAIN
B-Trees (generically, + or -) have very good lookup characteristics for random-access devices. They first found wide-spread use in relational databases something like 20 years ago. Basically the idea is to use a B-Tree with a very large radix. In the case of HAMMER, I am using a radix of 64 and could go all the way to 256 (the maximum that would fit in a 16K HAMMER buffer) if I wanted to. Use of a large radix greatly reduces the number of discrete I/O's (and ...
[ message continues ]
Jun 16, 7:53 am 2008
Thomas Zander
Re: HAMMER Update 13-June-2008 - HEADS UP: MEDIA CHANGED AGAIN
On Mon, Jun 16, 2008 at 10:28 AM, Matthew Dillon Maybe the question is why B+ trees are so popular in current file systems. I am certainly not expert enough to see all the implications of the selection of the tree structure on the implementation and performance of a file system, but what is the main reason why you chose B- trees? Riggs
[ message continues ]
Jun 15, 9:50 pm 2008
Dan M
Re: HAMMER Update 13-June-2008 - HEADS UP: MEDIA CHANGED AGAIN
On Sun, Jun 15, 2008 at 10:28 PM, Matthew Dillon They used B+ in earlier versions. In Reiser4 they decided to use http://en.wikipedia.org/wiki/B*-tree and http://en.wikipedia.org/wiki/Dancing_trees -- Dan
[ message continues ]
Jun 16, 1:18 pm 2008
Matthew Dillon
Re: HAMMER Update 13-June-2008 - HEADS UP: MEDIA CHANGED AGAIN
I am not too familiar with Reiser so I can't really come to that conclusion, but it doesn't seem likely that the reasons are similar. HAMMER's issue insofar as the B-Tree goes is mainly due to its history retention practices. If I mount with -o nohistory then the issue becomes one of locality of reference due to HAMMER not immediately reusing space freed by the rename-over that blogbench does. After running the test over the weekend the culprit seems to ...
[ message continues ]
Jun 15, 7:28 pm 2008
Aggelos Economopoulos
Re: tcpcb (was Re: sockbuf (was Re: BGL-free net stack))
What I've got seems to work (I can't easily test the ip_mroute changes), even though I'm not completely happy with the implementation. You (I don't mean just Sephe here) can get the changes by pulling the tcpcb branch from the git repo at git://repo.or.cz/dragonfly/netmp.git or by fetching http://leaf.dragonflybsd.org/~aggelos/B1-sopt.patch and applying it to HEAD. Please test in your regular setup and report any errors ("works for me" is interesting too). This patch does NOT contain any sockbuf ...
[ message continues ]
Jun 16, 10:16 am 2008
previous daytodaynext day
June 15, 2008June 16, 2008June 17, 2008