idea for concurrent seqscans

I had an idea that might improve parallel seqscans on the same relation.

If you have lots of concurrent seqscans going on a large relation, the
cache hit ratio is very low. But, if the seqscans are concurrent on the
same relation, there may be something to gain by starting a seqscan near
the page being accessed by an already-in-progress seqscan, and wrapping
back around to that start location. That would make some use of the
shared buffers, which would otherwise just be cache pollution.

I made a proof-of-concept implementation, which is entirely in heapam.c,
except for one addition to the HeapScanDesc struct in relscan.h. It is
not at all up to production quality; there are things I know that need
to be addressed. Basically, I just modified heapam.c to be able to start
at any page in the relation. Then, every time it reads a new page, I
have it mark the relation's oid and the page number in a shared mem
segment. Everytime a new scan is started, it reads the shared mem
segment, and if the relation's oid matches, it starts the scan at the
page number it found in the shared memory. Otherwise, it starts the scan
at 0.

There are a couple obvious issues, one is that my whole implementation
doesn't account for reverse scans at all (since initscan doesn't know
what direction the scan will move in), but that shouldn't be a major
problem since at worst it will be the current behavior (aside: can
someone tell me how to force reverse scans so I can test that better?).
Another is that there's a race condition with the shared mem, and that's
out of pure laziness on my part.

This method is really only effective at all if there is a significant
amount of disk i/o. If it's pulling the data from O/S buffers the
various scans will diverge too much and not be using eachother's shared
buffers.

I tested with shared_buffers=500 and all stats on. I used 60 threads
performing 30 seqscans each in my script ssf.rb (I refer to my
modification as "sequential scan follower" or ssf).

Here are some results with my modifications:
$ time ./ssf.rb # my script

real 4m22.476s
user 0m0.389s
sys 0m0.186s

test=# select relpages from pg_class where relname='test_ssf';
relpages
----------
1667
(1 row)

test=# select count(*) from test_ssf;
count
--------
200000
(1 row)

test=# select pg_stat_get_blocks_hit(17232) as hit,
pg_stat_get_blocks_fetched(17232) as total;
hit | total
--------+---------
971503 | 3353963
(1 row)

Or, approx. 29% cache hit.

Here are the results without my modifications:

test=# select relpages from pg_class where relname='test_ssf';
relpages
----------
1667
(1 row)

test=# select count(*) from test_ssf;
count
--------
200000
(1 row)

test=# select pg_stat_get_blocks_hit(17231) as hit,
pg_stat_get_blocks_fetched(17231) as total;
hit | total
--------+---------
199999 | 3353963
(1 row)

Or, approx. 6% cache hit. Note: the oid is different, because I have two
seperately initdb'd data directories, one for the modified version, one
for the unmodified 8.0.0.

This is the first time I've really modified the PG source code to do
anything that looked promising, so this is more of a question than
anything else. Is it promising? Is this a potentially good approach? I'm
happy to post more test data and more documentation, and I'd also be
happy to bring the code to production quality. However, before I spend
too much more time on that, I'd like to get a general response from a
3rd party to let me know if I'm off base.

Regards,
Jeff Davis