|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: iSCSI: MarkersJohn, FIM works now and COWS is not as bad as you paint it - as most software stacks touch the data at least once (no exceptions!). As for the pipeline argument - it must have something to do with Alaska - as I don't get it :-) If you want a vote - I can't - my heart is thorn ... Julo
OK, Julian, I buy that argument. I think that you have made the point that the value of Markers is Greater then I was giving credit for, in iSCSI/TOE integrated HBAs that are operating in CRC mode. That is great, since I am afraid we will not see the Framing stuff for some time. So the questions are which form of markers is the best. My vote comes down on the side of Fixed Interval Markers (FIM). For the following reasons: On the SW sending side, FIM works with low overhead whether CRC is being used or not. COWS, is a dramatic overhead unless the SW is performing CRC computation anyway. The COWS has a built in conflict with the Pipeline HW HBA, that has two different needs regarding the direction of the Pointers when talking to another partner HW HBA that uses a Pipeline approach. . . . John L. Hufferd Senior Technical Staff Member (STSM) IBM/SSG San Jose Ca Main Office (408) 256-0403, Tie: 276-0403, eFax: (408) 904-4688 Home Office (408) 997-6136, Cell: (408) 499-9702 Internet address: hufferd@us.ibm.com Sent by: owner-ips@ece.cmu.edu To: ips@ece.cmu.edu
Julian, There are two different issues with regards to CRC. One is on the sender side, and if Markers are done in Software, of course it matters if CRC is done, since if the sender does not use CRC, the cost of COWS is very high. The second point is about the Receiving Side. I did not understand your point, so I would appreciate you expanding on your comment that the Target Side would still get value out of Markers even if they use CRC. I can understand that there would be some additional value, especially in TCP Segment errors. By permitting some additional TCP Segments to arrive during the retry, but most of the value is lost because the Full Reassembly Buffer is needed for CRC. So now that I confess, that I do not completely understand your point, perhaps you could step me through your view of the value in the presents of CRC on a HW HBA/Chip. I think it is important that we are able to quantify the value. With regards to iWARP, I was using that name as an alias for the technique explained in the framing Draft (at http://www.ietf.org/internet-drafts/draft-ietf-tsvwg-tcp-ulp-frame-01.txt). But except for the last part of COWS, there is little resemblance between that and COWS. The framing does NOT use Word replacement. They only have a 8 byte header in common. That brings me to another point. If FIM needs to double its Markers (two words instead of one), would not COWS also need to double its Header Markers, for the same reason? . . . John L. Hufferd Senior Technical Staff Member (STSM) IBM/SSG San Jose Ca Main Office (408) 256-0403, Tie: 276-0403, eFax: (408) 904-4688 Home Office (408) 997-6136, Cell: (408) 499-9702 Internet address: hufferd@us.ibm.com Sent by: owner-ips@ece.cmu.edu To: ips@ece.cmu.edu cc: Subject: Re: iSCSI: Markers Some comments in text - Julo
When we were in Salt Lake City, we decided to hold a discussion on the Reflector regarding Markers. I think it is probably time to hold that discussion. Therefore, I will start it out. First I want to thank Julian for going to the trouble to document the proposal for COWS. This now gives us two Marker proposals to consider. They are the Fixed Interval Markers (FIM) and the Constant Overhead Word Stuffing (COWS) proposals. Having said that, lets examine each one and determine their usefulness. 1. Markers (FIM or COWS) are only needed by Hardware HBAs that are attempting to limit the amount of Reassembly Buffers they need on-board. This will always be a probabilistic determination by the vendors, but the purpose of Markers are to hold down the amount of total RAM needed, especially when factored into a probabilistic determination. 2. The use of Markers are negotiated per connection, and per direction. It is therefore possible for a software implementation to send markers, but not have to accept them. +++ that is specially important for software implementations ++ 3. When CRC Digest are negotiated to be used, it does no good to send out Markers since the hardware side will have to have Reassembly Buffers to compute the CRC Digest. Therefore, Markers do not help anything when CRCs are used on a connection. +++ That is not correct - Framing mechanisms are used to avoid an RTT related size for the receive buffer not avoinding a one-PDU buffer. Both FIM and COWS will help reducing HBA ememory requirements even when using CRC. When Using CRC a PDU reassembly buffer is unavoidale but it can be limited by the MaxRcvPDU. +++ 4. FIM can be implemented easily in software, and with an out going Scatter/Gather technique could send PDUs across a network to a Hardware HBA destination without requiring additional moves or touching each byte. 5. Software iSCSI implementations should always negotiate NOT to receive markers since Markers do not help the SW in any way. 6. Hardware HBAs can also implement FIM easily, and HW can do it easily in both directions. Therefore, if FIM is the approved Marker proposal, a HW HBA should send FIM whenever it sends iSCSI Commands and/or Data to other HW HBAs (assuming iWARP is not an available option), and the other HW HBAs wants them. 7. I do not foresee many installations using CRC Digests with Laptops and Desktops, since the overhead will probably be noticeable, and most installations do not judge Desktop and Laptop data to be key corporate assets (I am not saying that opinion is right, or that it is Universal, just that is a prevalent opinion). +++CRCs are not relevant +++ 8. A single software node would not be fast enough to cause a significant load or a significant memory requirement for the Reassembly Buffers. However, the combination of many (perhaps Desktops and Laptops spread across a real or virtual campus) can bring on a large load and the need for many Reassembly Buffers. The resultant amount of small Reassembly Buffers could make the Total requirement very High (thus raising the cost of the HBA). 9. Given the need for Desktops and Laptops to avoid noticeable degradation, and the probability that they will not be using CRC, means that the use of FIM is a compatible option when working across a network from a SW Node to an iSCSI HW HBA implemented Storage Controller. 10. COWS was also designed to be easily implemented in Software and in Hardware. 11. COWS will require the iSCSI Software implementations to touch each byte, as it looks for matches to its Framing Pattern within the PDUs. +++ except when implemented within the CRC, checksum where no ADDITIONAL touch is involved +++ 12. When CRC-32 digests are being computed, the Software will have to touch ever byte anyway, so the additional overhead to look for matches to the Framing Pattern is negligible. On the other hand FIM also works well with SW CRC-32 Digest computation. 13. But the premies I set above (which is clearly up to debate), is that Desktops and Laptops will not usually compute the CRC Digest. 14. COWS has the ability to have its pointing Markers, which point to the Framing Pattern Matches, point either forward or backwards. When being processed by software, the direction does not matter. So since the Markers should only be used on outgoing PDUs, the approprate direction of the pointers is what ever the destination needs. 15. Many vendors, especially at the higher speeds, are implementing a pipe line design. The pipeline receive process, will most likely want the COWS pointers to be forward pointing (pointing in the direction of the end of the PDU). Since the SW does not care, forward pointers should be fine. 16. If the HW Pipelining iSCSI HBA is to send COWS Markers to another HW Pipelining HBA, there may be a problem. The outgoing pipeline would prefer to have the pointers pointing backwards, but the receiving pipeline would prefer to have the pointers going forwards. Therefore with COWS, either the sending or the receiving HW Pipelining HBA will be unhappy and pipeline design will get very much more complicated. 17. Everyone likes iWARP better, but it requires changes to the Host SW TCP/IP stacks in the case of SW Initiators. Since these SW installations are probably going to be Desktops and Laptops, these systems will very likely have some non current version of the OS, (such as windows 2k, ME, etc.). Hence it is unlikely that they will have iWARP in the majority of the SW TCP/IP Stacks). +++iWARP needs framing as well and I would guess the technique will be close to what COWS is. Future convergence speaks for COWS+++ 18. If, for some reason, a Server uses Software iSCSI, it will probably have the latest and greatest OS software, and if that is true, and if iWARP is approved by the IETF then it will probably have the TCP/IP extensions that are required by iWARP. It will therefore use iWARP instead of any Marker solution. 19. If the Server uses Software iSCSI, and if iWARP is not approved by the IETF, then COWS or FIM are both reasonable choices for the Server, if the Server also uses the CRC option. If CRC is not used on the Server, FIM is a better match with the Server SW iSCSI. 19. iWARP is probably going to stay in IETF "experimental mode" for some time into the future. And without iWARP being specified by iSCSI as -- at least, "MAY Implement" and "MAY use", the HW to HW mode will still need a Marker solution. 20. Therefore, we are left with Markers as the only technique we can count on to aid and assist the iSCSI HW HBAs, keep their "on-board" RAM requirements to a minimum. Based on the above I come to the following conclusions: 1. Markers are needed. 2. FIM works best with the probable SW environment, and has the lowest overhead (since they will probably not use CRC Digests). +++CRCs are not relevant+++ 3. Since if CRC is used, the HW needs its Reassembly Buffers anyway, so there is no reason to use COWS, or any type of Marker. +++ again the same confusion betwen the PDU buffer and the RTT related holding are in the absence of Markers+++ 4. FIM can work well between HW HBAs without significant impact to pipeline design. +++ Somebody already said that FIMs are like democracy - imperfect, ugly but work+++ +++ However COWS are not that much heavier if carefully implemented and might get us faster to converge on a framing model+++ Therefore, I recommend that the Draft continue to define FIM. Also, I recommend that the FIM type of Markers be "MUST Implement" on the Out-Going direction, and "MAY Implement" on the incoming direction". Further, I recommend the Draft say that FIM "MAY be used" in any direction, depending on the negotiations. . . . John L. Hufferd Senior Technical Staff Member (STSM) IBM/SSG San Jose Ca Main Office (408) 256-0403, Tie: 276-0403, eFax: (408) 904-4688 Home Office (408) 997-6136, Cell: (408) 499-9702 Internet address: hufferd@us.ibm.com
Home Last updated: Tue Jan 08 13:17:51 2002 8309 messages in chronological order |