Re: Quick-and-dirty compression for WAL backup blocks

Hi all,

I examined the effect of block-hole compression patch.

I compared the amounts of writing of the WAL data of CVS(7/11) and
8.0.3. (The amount of the WAL data writing was measured by the number of
executions of the write() function in XLogWrite().)
And, I measured the size of the hole.

Environment;
IBM x206 P4 3.0GHz Mem 4GB
CentOS 4.0 (Linux 2.6.9-5.0.3.ELsmp)

Parameters;
shared_buffers = 65535
checkpoint_segments = 30
default_with_oids = false (8.0.3)
default_with_oids = off (CVS)

How to exam;
0) initdb --no-locale
1) pgbench -i -s 100 pgbench
2) pgbench -n -c 50 -t 5000 pgbench
3) vacuumdb -d pgbench
4) pgbench -n -c 50 -t 5000 pgbench

Results;
| 8.0.3 | CVS(7/11)
Exam | | | | | block-hole (byte)
| write |C.P| write |C.P| total | min | max | avg
-----+---------+---+---------+---+-----------+-----+------+---------
1) | 187505 | 3 | 187373 | 4 | 194056 | 36 | 8124 | 3881.12
2) | 509725 | 6 | 513489 | 5 | 115564476 | 12 | 8096 | 347.69
3) | 280456 | 2 | 172973 | 2 | 95923360 | 248 | 8156 | 614.08
4) | 533971 | 7 | 525135 | 6 | 171147256 | 12 | 8140 | 482.11

C.P = Checkpoint frequency

It has been understood that patchs seems to be effective at VACUUM as a
result of the measurement. But, in other cases, the effect was not so seen.

--
Junji Teramoto