[Xapian-discuss] Xapian on SSD vs SATA

[James Aylett]

On Fri, Oct 23, 2009 at 02:11:12PM +0200, Henry wrote:

> Using an SSD seems to only really (significantly) benefit phrase  
> searches (~10x faster compared to a HD).

This isn't surprising; for most searching, you're hitting the posting
list (postlist.DB in flint/chert). In a running system, the most
important bits of this will happily sit in cache so, as Richard has
pointed out, you probably aren't IO bound at this point. (Even with a
postlist somewhat larger than available memory, the most important
bits should still fit in core.)

However phrase searches hit the position table as well, which probably
won't be in cache. At this point, you'll be IO bound. You may remain
IO bound on subsequent phrase searches if enough of your other load
pushes the position table out of cache again quickly; or the important
parts of it plus the postlist won't fit into cache in the first place;
or if the phrase searches are hitting a wide spread of terms without
much commonality (since you'll keep on loading new bits from disk
until the whole lot is in there, or something in the VM gets smart and
prefetches the rest of it).

J

-- 
  James Aylett

  talktorex.co.uk - xapian.org - uncertaintydivision.org