DragonFly BSD
DragonFly kernel List (threaded) for 2010-02
[Date Prev][Date Next]  [Thread Prev][Thread Next]  [Date Index][Thread Index]

Re: kernel work week of 3-Feb-2010 HEADS UP

From: Michael Neumann <michaelneuma@xxxxxxxxxxxxxx>
Date: Fri, 5 Feb 2010 02:04:25 +0100

2010/2/5 Matthew Dillon <dillon@apollo.backplane.com>

:...but you are leaving some room for also using it as a write cache later,
:right?  Whether it's some special hook for HAMMER's un/redo log, or
:something completely filesystem agnostic, it seems it could be a huge
:benefit to some applications. (like databases?)
:Just thinking out loud: if you have some resident-like way of "locking"
:data on the SSD, that might be an interesting way to get something similar
:to ReadyBoost.  (That would perhaps make for a neat GSoC project?)

   Yes, it's theoretically possible by creating a small partition
   on the SSD (separate from the swap partition) and assigning
   all of HAMMER's UNDO FIFO disk blocks to it.  That would make
   fsync() considerably faster (65uS instead of 4-10ms).  It's
   not on my personal TODO list though.

Btw, that should be already possible using multiple HAMMER volumes. 
The SSD partition should be made the root volume. It can be very 
small, so that only the UNDO log fits on it (maybe a GB?), the second volume
would then be the regular hard disk. Maybe we'd need to give newfs_hammer
a specific option so that it treats all space of the first volume as UNDO and
uses the second volume for storage. 

I think I can implement that. Matt, do you think the option to newfs_hammer is
a good idea?



[Date Prev][Date Next]  [Thread Prev][Thread Next]  [Date Index][Thread Index]