Hello,

As blocksizes, random I/O and linear I/O are critical I/O performance parameters I had a look on PostgreSQL and a commercial software vendor.

Therefore I enhanced the system tap script: http://www.wiesinger.com/opensource/systemtap/disktop_gw.stp

Output per 5 seconds on a sequence scan:
    UID      PID     PPID                       CMD   DEVICE    T        BYTES  
   REQUESTS    BYTES/REQ
     26     4263     4166                postmaster     dm-1    R    168542208  
      20574         8192
=> 32MB/s

So I saw, that even on sequential reads (and also on bitmap heap scan acces) PostgreSQL uses only 8k blocks. I think that's a major I/O bottleneck.

A commercial software database vendor solved the problem by reading multiple continuous blocks by multiple 8k blocks up to a maximum threshold. Output per 5 seconds on an equivalent "sequence scan":
    UID      PID     PPID                       CMD   DEVICE    T        BYTES  
   REQUESTS    BYTES/REQ
   1001     5381        1                   process     dm-1    R    277754638  
       2338       118800
=> 53 MB/s

A google research has shown that Gregory Stark already worked on that issue (see references below) but as far as I saw only on bitmap heap scans.

I think this is one of the most critical performance showstopper of PostgreSQL on the I/O side.

What's the current status of the patch of Gregory Stark? Any timeframes to integrate? Does it also work for sequence scans? Any plans for a generic "multi block read count" solution?

Any comments?

Thnx.

Ciao,
Gerhard

--
http://www.wiesinger.com/

http://wiki.postgresql.org/wiki/Todo#Concurrent_Use_of_Resources
http://archives.postgresql.org/pgsql-hackers/2007-12/msg00027.php
http://archives.postgresql.org/pgsql-hackers/2007-12/msg00395.php
http://archives.postgresql.org/pgsql-hackers/2007-12/msg00088.php
http://archives.postgresql.org/pgsql-hackers/2007-12/msg00092.php
http://archives.postgresql.org/pgsql-hackers/2007-12/msg00098.php

http://archives.postgresql.org/pgsql-hackers/2006-10/msg00820.php

http://markmail.org/message/a5osy4qptxk2jgu3#query:+page:1+mid:hz7uzhwxtkbzncy2+state:results
http://markmail.org/message/a5osy4qptxk2jgu3#query:+page:1+mid:a5osy4qptxk2jgu3+state:results

--
Sent via pgsql-general mailing list (pgsql-general@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-general

Reply via email to