|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] RE: profiles - a way to simplify iSCSI> -----Original Message----- > From: owner-ips@ece.cmu.edu [mailto:owner-ips@ece.cmu.edu]On Behalf Of > Matt Wakeley > Sent: Friday, June 22, 2001 1:57 PM > To: ips@ece.cmu.edu > Subject: Re: profiles - a way to simplify iSCSI > > > The whole purpose for creating iSCSI in the first place (by the group that > started it) was to create a storage interconnect that could take > advantage of > the very high performance promised by 10GBE, at low cost. And share a common infrastructure (including common adapters) with all other data center apps such as NAS, http etc. > Mutliple > connections per session is there to take advantage of the 10GBE > link with the > shortcomings of TCP, How? By skipping around the congestion control algorithms? What about the other flows in the network? > and markers are there to eliminate the cost of memory > subsystems required to buffer out of order TCP frames. Framing has other benefit but not having a fast memory system cannot be avoided (especially if the inrastructure is to be common.) Also markers and CRCs do not mix well together. > Eliminate > these, and > you've you'll capitulate to Fibre Channel for anything but slow storage > connects. I don't think so. The pull of a common infrastructure will be too strong. And in case you have not looked lately, memory is getting fast enough and cheap enough. The GigaHz PCs are driving this. Somesh > > -Matt > > John Vrabel wrote: > > > > > It's time to start thinking about what we can take out ... > > > > I fully agree with the need to limit options, and avoid the need for > > profiles. > > > > FWIW, our list of things to remove: > > > > 1) Markers > > > > 2) SNACKs > > > > 3) Command retries > > > > 4) Async Messages > > > > 5) Multiple connections per session > > > > Things to change: > > > > 1) default number of connections from 8 to 1 ( if multiple > connections > > remain ) > > > > Things I'm less sure about: > > > > 1) InitialR2T, ImmediateData. Do they really need to be > negotiable? > > FirstBurstSize controls them to a great extent. > > Can InitialR2T be removed, and retain ImmediateData for short > > writes ? > > > > John Vrabel _________________________________________________________ Do You Yahoo!? Get your free @yahoo.com address at http://mail.yahoo.com
Home Last updated: Tue Sep 04 01:04:24 2001 6315 messages in chronological order |