Thread: Results of testing WAL

I've run some tests with 7.1 + WAL & 7.0.2
Setup: 5 tables (i int, t text), 100000 records in each table,
sizeof(t column) is rand(256), indices on i column for all tables.
-B 16384 -A 0 (+ --wal_buffers=256 in 7.1)

System: SUN Ultra 10, 512M RAM, 1 (fast) IDE disk

Test: 5 clients simultaneously performed
UPDATE tableN SET t = '...rand(256) chars...' WHERE i = ...rand(100000)...;
Each UPDATE was in separate transaction, client N changed tableN only,
each client made 1000 transactions.

Results: 5000 transactions took ~60 sec in 7.1, ~550 sec in 7.0.2 with fsync
and ~60 sec without fsync.

So, seems that WAL added not just complexity to system -:)

I'm going to commit redo for sequences tomorrow evening and
#define XLOG by default after this (initdb will be required).

Vadim

"Mikheev, Vadim" <vmikheev@SECTORBASE.COM> writes:
> Results: 5000 transactions took ~60 sec in 7.1, ~550 sec in 7.0.2 with fsync
> and ~60 sec without fsync.

> So, seems that WAL added not just complexity to system -:)

Sounds great!

> I'm going to commit redo for sequences tomorrow evening and
> #define XLOG by default after this (initdb will be required).

I suggest bumping the catversion.h number when you #define XLOG,
so that people won't be able to accidentally start an old postmaster
with new DB or vice versa.
        regards, tom lane

> > I'm going to commit redo for sequences tomorrow evening and
> > #define XLOG by default after this (initdb will be required).
> 
> I suggest bumping the catversion.h number when you #define XLOG,
> so that people won't be able to accidentally start an old postmaster
> with new DB or vice versa.

Thanks!

Vadim

> > Results: 5000 transactions took ~60 sec in 7.1, ~550 sec in 
> > 7.0.2 with fsync and ~60 sec without fsync.
> > 
> > So, seems that WAL added not just complexity to system -:)
> 
> Wow, this sounds fantastic :-) 
> I see my concerns where not justified.

Let's see first how justified are my hopes that WAL code
are bug free -:)

Vadim