|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] RE: ISCSI: Urgent Flag requirement violates TCP.Dante, All you say makes sense except that the sequence number won't help you the iSCSI packet header gets lost. Then you have to either be able to find fast the next one (and that what the urgent pointer and bit are here for) or pray... This is all stuff that was repeated several time on this reflector and a quick search in the Archive helps always. Julo "Dante Malagrino'" <dantem@cisco.com> on 14/11/2000 20:32:34 Please respond to "Dante Malagrino'" <dantem@cisco.com> To: "Y P Cheng" <ycheng@advansys.com> cc: ips@ece.cmu.edu Subject: RE: ISCSI: Urgent Flag requirement violates TCP. At 09:18 AM 11/14/2000 -0800, Y P Cheng wrote: >Well, the reasons for continuing to move data as fast as the wire speed in >the face of a sequence hole are: 1) out-of-order reception is considered >normal and happens often, 2) On a ten gigabit backbone, several milliseconds >of delay requires the buffering of several megabytes of data on an adapter. >(One megabyte per millisecond). Incoming data are not limited from a single >source. Many nodes may send and return data to an adapter at the same time. >At a gigabyte per second incoming rate, using several megabytes of SRAM for >buffering on an adapter is very expensive. 1) what do you mean by "out-of-order reception is considered normal and happens often"? Last figures I saw talked about 3% of out-of-order TCP segments on WANs (worst-case). Is it a critical value? 2) with 64MB of RAM, and not supporting Window Scale Option, given a maximum Window Size of 64k (I have seen it very rarely), we end up having 1k-conns concurrently communicating with a given adapter. With a more reasonable Window Size (16k), the number of connections is four times more. Do we expect to see more connections than that? Do we plan to support Window Scale Option? >If there is any possibility for a TOE adapter to learn the beginning of an >iSCSI PDU in face of a sequence hole, while it could try to keep in-order >delivery of command and status PDUs -- although may not be necessary, the >data PDUs can be moved quickly to the buffers pre-allocated by application >software. Hence, it will greatly reduce the buffering requirement of the >adapter. TCP is bytestream oriented, and each packet carries a Sequence Number. Why can't you save an out-of-order packet in his appropriate location (in the application buffer), and then put the missing packet in the appropriate hole? There is no need for copy. >If a TOE adapter can't move at the speed of the wire, how are we going to >take advantage of the 10 Gbps media? On a large WAN with high speed >backbone connections with 100 millisecond latency, there could be 100 >megabytes of data inflight. Can we buffer all 100 megabytes on an adapter >or should we limit the inflight data by set a small TCP window limited by >the buffer size of the adapter? Even with large window sizes (64k), how do we have 100MB inflight? See my previous comment. >The right design of a TOE adapter is always >to move data quickly to the buffers already allocated by application >software and to allow as much data inflight as possible. To achieve that, >the TOE adapter needs all the help it can get. If we can't move data at the >wire speed, lets not bother to build 10 Gbps networks. Isn't the help provided by the TCP sequence number enough? -- Dante Dante Malagrino' Cisco Systems Empowering the Internet Generation 170 West Tasman Drive Tel. (408) 525 4120 San Jose, CA, 95134-1706 Fax. (408) 525 4120 dantem@cisco.com
Home Last updated: Tue Sep 04 01:06:24 2001 6315 messages in chronological order |