Re: Checkpoint cost, looks like it is WAL/CRC

I wrote:
> We still don't know enough about the situation to know what a solution
> might look like. Is the slowdown Josh is seeing due to the extra CPU
> cost of the CRCs, or the extra I/O cost, or excessive locking of the
> WAL-related data structures while we do this stuff, or ???. Need more
> data.

Josh, is OSDL up enough that you can try another comparison run?
If so, please undo the previous patch (which disabled page dumping
entirely) and instead try removing this block of code, starting
at about xlog.c line 620 in CVS tip:

/*
* Now add the backup block headers and data into the CRC
*/
for (i = 0; i < XLR_MAX_BKP_BLOCKS; i++)
{
if (dtbuf_bkp[i])
{
BkpBlock *bkpb = &(dtbuf_xlg[i]);
char *page;

COMP_CRC32(rdata_crc,
(char *) bkpb,
sizeof(BkpBlock));
page = (char *) BufferGetBlock(dtbuf[i]);
if (bkpb->hole_length == 0)
{
COMP_CRC32(rdata_crc,
page,
BLCKSZ);
}
else
{
/* must skip the hole */
COMP_CRC32(rdata_crc,
page,
bkpb->hole_offset);
COMP_CRC32(rdata_crc,
page + (bkpb->hole_offset + bkpb->hole_length),
BLCKSZ - (bkpb->hole_offset + bkpb->hole_length));
}
}
}

This will remove just the CRC calculation work associated with backed-up
pages. Note that any attempt to recover from the WAL will fail, but I
assume you don't need that for the purposes of the test run.

regards, tom lane