|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] RE: iSCSI: No FramingDe-lurking... Right now, I'd say disk-disk backup/restore and mirroring. The individual disk devices of today can perform at ~70 MB/s. One need only gang 14 of these together, streaming, to reach nearly 1000 MB/s, or 1 GB/s, or 8 Gb/s. The disk devices of tomorrow are going to perform at rates approaching 100 MB/s per spindle. Only ten devices are then necessary. This would be a single TCP stream (one initiator, one target, many LUs). If one has to mirror (/move) 10 TB in a reasonable amount of time, say, then a stream of 1 GB/s is not only necessary, but paramount. Even at that rate, it would still take nearly three hours to move just 10 TB. In a business-critical recovery situation, three hours is an eon. Don't think application (CPU/memory in a server) to storage. Think storage to storage. Rob Rob Peglar Corporate Architect XIOtech Corporation, a Seagate Company (314) 308-6983 >-----Original Message----- >From: Prasenjit Sarkar [mailto:psarkar@almaden.ibm.com] >Sent: Sunday, February 03, 2002 9:23 PM >To: THALER,PAT (A-Roseville,ex1) >Cc: ips@ece.cmu.edu; WENDT,JIM (HP-Roseville,ex1); >owner-ips@ece.cmu.edu >Subject: RE: iSCSI: No Framing > > > >Hi Pat, > >I had a few questions about your 10 Gbps example. If we assume >that this is >a single TCP stream, I would like to know which application in >the near and >forseeable future can sustain a 10 Gbps data rate for a >reasonable amount >of time? Or if they are multiple TCP streams, are you assuming >packet drops >in all streams simultaneously? > >Thanks, > > > Prasenjit Sarkar > Research Staff Member > IBM Almaden Research > San Jose > > > > > > "THALER,PAT > > (A-Roseville,ex1) To: >"WENDT,JIM (HP-Roseville,ex1)" > " ><jim_wendt@hp.com>, ips@ece.cmu.edu > <pat_thaler@agile cc: > > nt.com> Subject: RE: >iSCSI: No Framing > Sent by: > > owner-ips@ece.cmu > > .edu > > > > > > 02/03/2002 03:01 > > PM > > > > > > > > >Jim, > >In answer to your questions: > >Agilent is planning to implement framing. We view both FIM and >COWS as having about the same utility so we would implement >whichever one went into the standard. > >A smoothing buffer on the chip is feasible wrt your point 2. >There are some configurations that would use external memory. > >Also, one concern is that the very situation where one would >need large window size for efficiency - high bandwidth long >distance communication - is where out of order receipt becomes >increasingly likely so the amount of memory for this situation >could push up cost to an excessive extent. > >Reducing the amount of traffic that has to be shunted to an >external memory affects the bandwidth that needs to be provided >for that memory. If we can handle most PDUs internally then the >external memory doesn't have to be as wide and as fast. For >instance, if a drop means that the iSCSI PDU with the drop >is put into external memory then that takes less memory bandwidth >than if a drop means the whole data stream for that session will >be going into the buffer until the hole is filled. > >This decision also effects latency after a drop and required >backplane bandwidth. > >Let's say one is on a 10 Gbit/s extreme long distance WAN >connection and that there is a drop. If the round trip delay >for getting the whole filled is 100 ms, then without framing, >one could have about 1 Gbit stored in the local buffer memory >by the time the hole is filled. One will have to process all >of this through iSCSI and across the backplane before one can >deal with the new traffic coming in. If the backplane data rate >is closely sized to the external data rate, an extra 100 ms latency >has just been inserted into the path until traffic slacks off. >(Congestion control might mitigate this to the extent that one >is talking to just one source, but with multiple sessions, one >will still have to keep up.) To get the buffer emptied to make >space for further drops and to get the latency back down, one >will have to size the backplane bandwidth wider and have the >iSCSI engine process at a higher data rate. > >With FIM or COWS, one can end the incident with much less data >in the buffer as one processes PDUs after the hole while waiting >for it to be filled. Then, when the hole is filled one just has >to process a small amount of data to catch up. > >Regards, >Pat Thaler > >-----Original Message----- >From: WENDT,JIM (HP-Roseville,ex1) [mailto:jim_wendt@hp.com] >Sent: Tuesday, January 29, 2002 10:47 PM >To: ips@ece.cmu.edu >Subject: iSCSI: No Framing > > > >So, it would be good to hear from several iSCSI >NIC/chip implementors who: >- have or plan to implement FIM or COWS (or some >other framing mechanism) and take advantage of it to >minimize demands on on-NIC buffer memory >bandwidth/quantity. >- believe that all-buffers-on-chip solutions are >feasible and valid (wrt the points above, including >#2) >- can quantify the memory/pin/power/space cost >savings for all-buffers-on-chip solutions > >Jim > > > >
Home Last updated: Mon Feb 04 13:18:05 2002 8616 messages in chronological order |