One Byte is Seven bits too many? - A Design suggestion
Hi,
The norm takes up 1 byte of storage per document per field. While this may seem
very small, a simple calculation shows that the IndexSearcher can consume lots
of
memory when it caches the norms. Further, the current implementation loads up
the
norms in memory as soon as the segments gets loaded. Here are the calculations:
For Medium sized archives
docs=40Million, Fields=2 => 80MB memory
docs=40Million, Fields=10 => 400MB memory
docs=40Million, Fields=20 => 800MB memory
For larger sized archives
docs=400Million, Fields=2 => 800MB memory
docs=400Million, Fields=10 => ~4GB memory
docs=400Million, Fields=20 => ~8GB memory
To further compound the issues, we have found JVM performance drops when the
memory
that it manages increases.
While the storage itself may not be concern, the runtime memory requirement can
use
some optimization, especially for large number of fields.
The fields itself may fall in one of 3 categories
(a) Tokenized fields have huge variance in number of Tokens,
example - HTML page, Mail Body etc.
(b) Tokenized fields with very little variance in number of token,
example - HTML Page Title, Mail Subject etc.
(c) Fixed Tokenized Fields
example - Department, City, State etc.
The one byte usage is very applicable for (a) and not for (b) or (c). In
typical
usage, field increases can be attributed to (b) and (c).
Two solutions come to mind:
(1) If there is forethought in the field design, then one can prefix the field
tokens
and then reduce the number of fields. Of course, this will add the overhead of
embarrassing explanation to every query writer of why to add Prefix for every
token.
If however, this prefix can be done underneath, it may work but still not
elegant.
(2) The norm values for (c) has only two values. One is 0 when the field is
not present,
and the other value is a fixed one. In this scenario, the memory requirement
is only one bit per doc per field. I would argue that even for (b) one can
approximate the
value with one bit and not much loose much in ranking of documents.
Several implementation options are possible:
(a) We can implement the approximation at the time of writing index (Backward
compatibility
has to be considered)
(b) Use a bitset instead of an array for search purposes. I have been
wanting to do for the last 6 months, but have found time yet. If I do, will
submit an
implementation.
Also, if a field is mandatory, then the 0 scenario never occurs and in this
situation, we
use a single constant to represent the array. May be One byte is 8 bits too
many:-))
Arvind.
---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
