On Thu, Jan 31, 2002 at 08:39:34AM -0800, William Lee Irwin III wrote:
> For a branch factor of 128 on i386, this ends up as 1 + 7 + 7 + 6 = 21
> words per file. So for 500K inodes each with one page, 42MB (Douglas
> Adams fan?). Offsets of 10GB don't work here. Sounds like either an
> interesting patch or a 64-bit machine if they work for you. =)
As just pointed out to me the minute I did a substitution I went wrong:
A branch factor of 128 leads to
1 + (7 + 7 + 6)*128 words = 2561 words per file, which is somewhat
more severe. =(
More corrections are welcome.
On Thu, Jan 31, 2002 at 03:36:07PM +0100, Andrea Arcangeli wrote:
>> Otherwise for an unix fs developer usage (the small files ala dbench)
>> the rbtree was much nicer data structure than the hash in first place
>> (and it eats less ram than the radix tree if only one page is queued
>> etc...).
On Thu, Jan 31, 2002 at 08:39:34AM -0800, William Lee Irwin III wrote:
> And the pointer links in struct page? Sounds like more RAM to me...
> 4000 open files (much more realistic than 500K) each with one page
> leads to 48000 words of radix tree overhead. 3 words per page of
> pointer links and > 16000 pages of RAM and the rbtree eats more, not
> less. And 16000 pages is just 64MB on i386.
This doesn't quite hold up after the the correction above. 4K open
files ends up having 10.5K words or 40MB of overhead.
Cheers,
Bill
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/