|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] RE: ISCSI: Urgent pointer consensusI have in the past raised concern about it (coalescing of pointer). Others have raised a concern about how it may impact TCP accelerated adapters (URG pointer presence may kick the packet into non-fast path) - also software interface is non-efficient. It seems that if a connection is going at a rate where the data transfer rate is equal or higher than the producer rate of generating data, and there is a segment loss, then the urgent pointer is going to help place the data in a proper location (if the dropped PDU contained an iSCSI header and iSCSI PDU < window size or thereabouts). However, it is not clear to me what happens in the face of slow start and other exception conditions. I have some analysis below of what might happen if iSCSI implementation and TCP implementation are not literally mushed together. ---- First of all, slow start is likely to cause coalescing of the Urgent pointer (at least according to my interpretation of the 793) if the sender all of a sudden wants to send a lot of data or the rate of production is greater than the rate of transmission - even for short amounts of time. Matt suggested that in case of a segment loss, the adapter would look for the next urgent pointer and drop all the data between the lost segment and the next urgent data. The result of this behavior would actually cause the sender to make some assumptions about link congestion (even in the presence of SACK), the results of which are not likely to be good. Also, what happens when a lost segment is retransmitted? Is it supposed to carry the URG pointer, wherever it may point? ---- It seems fairly easy to create a model where iSCSI and TCP are closely tied together and make some assumptions about the "rate at which the application - iSCSI - is generating packets". In this case we can use the URG pointer to have the desired behavior - we should negotiate it i.e. make it optional. And also very clearly specify the behavior since RFC793 is ambigous (what happens when there is retransmission? how to distinguish that from out of order?). The behavior may include negotiated limits on the size of the iSCSI PDUs. Somesh > -----Original Message----- > From: Black_David@emc.com [mailto:Black_David@emc.com] > Sent: Tuesday, November 14, 2000 11:17 AM > To: ips@ece.cmu.edu > Subject: RE: ISCSI: Urgent pointer consensus > > > > > I also believe that based on discussion on the list, WG > rough consensus > > > does NOT exist for requiring this use of the URG flag and Urgent > pointer, > > > (too many people have objected to making them mandatory), > and hence > > > the current "MUST" will have to be replaced. > > > > Um, who, besides Doug, has objected? Seems to me most of > the discussion > has > > been centered around why it would be useful. Once the > reason(s) have been > > explained, I don't see too much dissention. > > That's not what I see on the list. For starters, Daniel Smith, > Glen Turner, and Ronald Lee have objected to the mechanism > or specifically to making it mandatory, and I think Silvano > Gai just added himself to that list. More people can probably > be found by searching the list archives. Please don't confuse > having the last word with winning the argument. > > Thanks, > --David > > --------------------------------------------------- > David L. Black, Senior Technologist > EMC Corporation, 42 South St., Hopkinton, MA 01748 > +1 (508) 435-1000 x75140 FAX: +1 (508) 497-8500 > black_david@emc.com Mobile: +1 (978) 394-7754 > --------------------------------------------------- >
Home Last updated: Tue Sep 04 01:06:25 2001 6315 messages in chronological order |