|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] RE: iSCSI: Framing StepsWhile I support a generic direct data placement model, the following are additional points for consideration for analysis of memory bandwidths and sizes. 1. A 10G link dropping packets will not do TCP at 10Gbps. The rate drops as the packet loss increases. I don't recall but Franco or Victor from Nortel had posted an equation once. 2. The amount of memory needed is of the order of delay-bandwidth product and is reasonably independent of the number of connections. Of course, statistically, worse situations are possible, but that is not a design point. If lots of packets are being lost, things are going significantly slower anyway (thus reducing the delay- bandwidth product). 3. The analysis of 10Gbps, half way round the world was initially used in this debate. ALthough interesting, the person with the scenario above is not going to blink at the cost of 256MBytes of fastest memory considering what they are paying for the link. 4. All assumptions are based on existing SCSI API models at targets. Changes to target APIs/meta-data management would let some vendors not have to do data copies anyway. 5. A solutions that is iSCSI specific is not likely to give the savings (especially when it may not work) people expect. As I mentioned, I think we should take the same approach as we did for error recovery (if it is not spilling the beans too much) and come up with a more detailed analysis. Somesh > -----Original Message----- > From: owner-ips@ece.cmu.edu [mailto:owner-ips@ece.cmu.edu]On Behalf Of > Stephen Bailey > Sent: Tuesday, January 29, 2002 6:43 AM > To: ips@ece.cmu.edu > Cc: digital_aryan@yahoo.com > Subject: Re: iSCSI: Framing Steps > > > Sridhar, > > > Its true that the current day IP networks are capable of handling packet > > drops .... But it is high time to have a mechanism to handle this > > more efficiently, keeping in view the importance of SANs and their > > intended use. > > The combination of IP+transport (TCP/IP) is precisely designed to > deliver data efficiently and scalably (as a function of network size). > It is the transport layer's role to maximize `goodput'. Dropping > packets in the network as a result of congestion and having the > transport adjust to this is what results in IP's scalability as > compared to system area (flow controlled) network technologies. In > other words, it's a feature, not a bug. It's a large part of what > made IP the dominant networking technology, and it's never going away. > > It is not the case that the designers of TCP said `heck, we've got > these miserable, lossy networks, I guess we'll just suck it up and do > the best we can, even though it'd work a lot better if we had some > other characteristics.' > > TCP does have some warts in providing the best possible goodput, but > those are problems for TCP (or other transport) to fix. Anyway, that > doesn't seem to be what you're referring to. > > The RFC 1122 SHOULD says TCP endpoints should cling tightly to each > correctly received byte. If an implementation is discarding portions > or entire flights of data because a packet is lost that's certainly > NOT TCP's problem, and not what FIM is intended to fix. FIM's (or any > framing technique's) ONLY goal is to reduce the COST (which may equate > with `tractability' if the cost is otherwise prohibitive) of an > efficient implementation. It does not improve the data transport > efficiency itself at all. > > Steph > > >
Home Last updated: Thu Jan 31 10:18:06 2002 8572 messages in chronological order |