Tom Lane <tgl@sss.pgh.pa.us> writes:
> Greg Stark <gsstark@mit.edu> writes:
> >>> What do you think is broken about fragmented UDP packets?
>
> > Fragmentation happens at the IP protocol level, the kernel is responsible for
> > reassembly. There's nothing for the application level to handle.
>
> And, by the same token, on platforms where it is broken there is nothing
> we can do about it.
Ok having read the source I think I see why we're talking past each other.
The pgstat messages are indeed fixed size. So bumping the size up to 8k would
mean *every* udp packet would be bumped up to 8k.
However there's no good reason for that to be the case. recv(2) always returns
exactly one packet and tells you how large it was. And in fact the
PgStat_MsgHdr even has a redundant message size field that could serve the
same purpose. So we actually have all the machinery needed twice over to avoid
the fixed size messages.
In fact looking over the code I think it would be simple to change this. I
think it would be reasonable to make just the PgStat_MsgActivity variable
sized. I'll look at it a bit more, I think it's well worth the slight code
complexity, especially if it were only done for the one message type.
This has potential to reduce the average size of these messages quite a lot.
Potentially reducing the data being pumped through udp and the pipe buffer
(?!) quite a lot.
--
greg
