[erlang-questions] File writing and odd behavior when unrelated IO activity occurs

[Witold Baryluk]

Dnia 2009-07-16, czw o godzinie 12:39 -0600, Rob Elsner pisze:
> All,
> 
> I'm recording a multicast stream using gen_udp in active mode (so my
> process receives a mailbox message every time a packet arrives).  This
> is for a 35 megabit/second stream.  I have a simple worker process
> doing the recording:
> 
> loop(FileHandle, Socket, MulticastAddress, MulticastPort,
> SourceAddress, FilePrefix) ->
> 	receive
> 		{udp, _, _, _, Bin} ->
> 			case file:write(FileHandle, Bin) of
> 				ok ->
> 					recorder:loop(FileHandle, Socket, MulticastAddress,
> MulticastPort, SourceAddress, FilePrefix);
> 				{error, Reason} ->
> 					close_file(FileHandle), % close our current file
> 					close_socket(Socket, MulticastAddress, SourceAddress),
> 					log4erl:error("file:write returned error ~p", [Reason])
> 			end;
> 		{mark} ->
> 			NewName = get_file_name(MulticastAddress, MulticastPort,
> SourceAddress, FilePrefix), % figure out the new file name based on
> address and time
> 			log4erl:info("Recording mark ~p ~p ~p New name ~p",
> [MulticastAddress, MulticastPort, SourceAddress, NewName]),
> 			{ok,NewHandle} = open_file(NewName), % open it
> 			close_file(FileHandle), % close our current file
> 			recorder:loop(NewHandle, Socket, MulticastAddress, MulticastPort,
> SourceAddress, FilePrefix); % keep writing!
> 		{stop} ->
> 			close_file(FileHandle), % close our current file
> 			close_socket(Socket, MulticastAddress, SourceAddress)
> 		after 10000 ->
> 			log4erl:warn("Timed out ~p ~p ~p
> ~p~n",[MulticastAddress,MulticastPort,SourceAddress,FilePrefix]),
> 			close_file(FileHandle), % close our current file
> 			close_socket(Socket, MulticastAddress, SourceAddress)
> 	end.
> 
> These files get segmented at around 13GB for an hour long recording.
> The issue I am seeing is that using the above process, if I delete a
> file which has already been recorded, it causes the file:write process
> to "hang" which causes my mailbox to fill up and consume all 8GB of
> RAM on this box, before crashing the VM with an OOM error.
> 
> I've tried the file open options with raw set and not set, with
> delayed_write.  Does anyone have any thoughts about why a file delete
> operation on an unrelated file would cause file:write to block or take
> an excessive amount of time? I've tried ionice'ing the erlang VM to
> realtime highest priority, to no avail.
> 
> Thanks,
> Rob Elsner
> 
> ________________________________________________________________
> erlang-questions mailing list. See http://www.erlang.org/faq.html
> erlang-questions (at) erlang.org
> 


Are you using ext3? There is problem on unlinking big files, data blocks
need to be deallocated, and on ext3 this process locks other operations.
I would recomend XFS on Linux for big files, especially in concurant
write/read scenario.



-- 
Witold Baryluk <baryluk@REDACTED>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 197 bytes
Desc: To jest cz??? wiadomo?ci podpisana cyfrowo
URL: <http://erlang.org/pipermail/erlang-questions/attachments/20090716/a34a96f6/attachment.bin>