tsearch2/GIST performance factors?

We are indexing about 5 million small documents using tsearch2/GIST. Each "document" contains 2 to 50 words. This is a "write once, read many" situation. Write performance is unimportant, and the database contents are static. (We build it offline.)

We're having problems with inconsistent performance, and it's very hard to separate the effects of various factors. Here are the things we think may be relevant.

1. Total number of words

Our documents currently contain about 110,000 unique words. Oleg wrote: "[The limit is] 100K, but it's very fuzzy limit." By trial and error, we've learned that 50,000 works well, and 150,000 works poorly, so Oleg's comment appears to be a good rule-of-thumb. (With SIGLENINT enlarged, see below.) But there may be other factors that affect this conclusion (such as shared memory, total memory, etc.).

2. Total size of the table

5 million documents is not a very big database (each document is a few to a few hundred bytes), so we don't think this is relevant.

3. Number of documents per word

There seems to be a VERY strong effect related to "common" words. When a word occurs in more than about 1% of the documents (say 50,000 to 150,000 documents), performance goes WAY down. Not just for that specific query, but it screws up tsearch2/GIST completely.

We have a test of 100 queries that return 382,000 documents total. The first time we run it, it's slow, about 20 minutes (as expected). The second time we run it, it's very fast, about 72 seconds -- very fast!! As long as we avoid queries with common words, performance is very good.

But, if we run just one query that contains a common word (a word that's in more than about 2% of the documents, roughly 150,000 documents), then the next time we run the 100 test queries, it will take 20 minutes again.

We can't simply eliminate these common words. First of all, they can be very significant. Second, it doesn't seem like 2% is "common". I can understand that a words like "the" which occur in most documents shouldn't be indexed. But a word that occurs in 2% of the database seems like a very good word to index, yet it causes us great problems.

I've read a bit about tsearchd, and wonder if it would solve our problem. For our application, consistent performance is VERY important. If we could lock the GIST index into memory, I think it would fix our problem.

I tried copying the GIST indexes (which are in a separate tablespace) to a 1 GB RAM disk, and it made the initial query faster, but overall performance seemed worse, probably because the RAM disk was using memory that could have been used by the file-system cache.

4. Available RAM and Disk drives

Would more RAM help? How would we tell Postgres to use it effectively? The GIST indexes are currently about 2.6 GB on the disk.

Would more disks help? I know they would make it faster -- the 20-minute initial query would be reduce with a RAID drive, etc. But I'm not concerned about the 20-minute initial query, I'm concerned about keeping the system in that super-fast state where the GIST indexes are all in memory.

Hardware:
Dual-CPU Xeon Dell server with 4 GB memory and a single SATA 7200 RPM 150GB disk.

tsearch2/gistidx.h
modified as: #define SIGLENINT 120

System configuration:
echo 2147483648 >/proc/sys/kernel/shmmax
echo 4096 >/proc/sys/kernel/shmmni
echo 2097152 >/proc/sys/kernel/shmall

Postgres Configuration:
shared_buffers = 20000
work_mem = 32768
effective_cache_size = 300000

I feel like I'm shooting in the dark -- Linux, Postgres and tsearch2/GIST are interacting in ways that I can't predict or analyze. Thanks very much for any comments and advice.

Craig