[pnfs] notes on todays tele-con

[William A.(Andy) Adamson]

thought it might be useful to distribute my notes...

-->Andy


-------------- next part --------------
Bakeathon
---------

Ordered TODO list
1) debug 2.6.17 client.
2) large I/O patch from greg banks added to vendor base in cvs
3) finish all operations 
	Client layoutreturn  -  Raul
	Client and Server layout recall -  mark, andy (server)
4) move to latest draft
5) remove GETATTR from data server i/o
5) client recovery - fall back to MDS
	 	   - if ds fails, look at device list for another server.
6) RPCSEC_GSS kerberos to data servers and mds (?)

bakeathon testing
	- pynfs tests ??
	- server functionality for admin to recalls layouts - use for testing
	- recovery from data server loss (GPFS only)
	- direct I/O testing with filelayout
	- need more testing ideas	

Beyond bakeathon
	- layout cache on client
	- what is the policy for layoutget after a recall?
	- sub-dividable layouts (required for objs/blocks)
	- pNFS on (basic) sessions (porting existing code forward)

Bakeathon pNFS implementations:
	- Linux client
	- Sun might have a client. regular read/write functionality only
	- GPFS, Network Appliance, Panasas, DESY/dCache back ends
	- other?

General Notes
-------------

layoutreturn:
	need arguments but generic layout is opaque. 
	inode gets reaped - free layout structure, return layout.
	for now assume full file layout

object based layout limited in time, different from v4 lease.
	- put a barrier on layout. wait for outstanding I/O.
	- layout commit prior to layoutreturn
	- what is the scope of the barrier lock
block based layout:
	- reclaim bit on layouts for server reboot
	- modified cached block layout needs to sync with server after server
	  reboot.