Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

RE: [RPRWG] Scheduling



Dear Wolfram,
 
    I do agree with your analysis, and like to amplify
two points you made (and you may or may not agree that
these two are as important as I make it to be).
 
1. We are not standardizing overall QoS architecture.  What
   we are building is fair-access method of the shared medium
   ring MAC that is CoS-aware.  All discussions of DiffServ/802.1P
   class definition is applicable to queue on the buffer-insertion
   path, not to the fair-access aspect of the shared ring access.
 
   My observation is that the vocal part of this reflector is
   converging to 3 CoS awareness, with very specific behavior of
   each.  If enough discussion has occurred on this matter, we
   should be able to vote on this on our next meeting.  If you
   feel otherwise, speak up (or speak up further).
 
2. Store-and-Forward, versus Cut-Through, may be a non-issue for
   high-speed (e.g. >=2.5G) network to meet jitter/delay, etc.
   requirements for non-data (e.g. TDM, VoIP) payload. 
 
   My preference on this matter is to allow for both, so long as
   interoperability is not compromised for MACs >=1G.  We do this
   by specifying max. latency specification for a transit frame
   when ring input had been in idle.  So it is tied to the objectives
   of specifying Ring MAC >=1G. 
 
   15xx byte Store-and-Forward provides 3.1 mS latency for 256 node
   ring at 1G w/o fiber link length delay.  256 node, with 40 Km
   between any two adjacent nodes provide latency of
       (40 Km/link / (250 m/uSec))*256 links = 41 mS.
   So from latency point of view, unless the network is provisioned,
   it would not meet latency of <= 10 mS for voice type of payload.
 
   Jitter performance will be different between the two.  But since
   any assumption I make in this area is expected to be controversial,
   I'll leave it up to the advocates of each camp.  But I would still
   like to pursue max. jitter specification within the context of the
   fair access mechanism discussed in #1.  The jitter performance is
   a gross function of the two, fair-access MAC behavior and transit
   buffer behavior.
 
regards,
 
Yong.
 
============================================
Yongbum "Yong" Kim      Direct (408)922-7502
Technical Director      Mobile (408)887-1058
3151 Zanker Road        Fax    (408)922-7530
San Jose, CA 95134      Main   (408)501-7800
ybkim@xxxxxxxxxxxx      www.broadcom.com
============================================
 
-----Original Message-----
From: owner-stds-802-17@xxxxxxxx [mailto:owner-stds-802-17@xxxxxxxx]On Behalf Of Harmen van As
Sent: Thursday, April 05, 2001 6:05 AM
To: stds-802-17@xxxxxxxx
Cc: Wolfram Lemppenau
Subject: RE: [RPRWG] Scheduling

Based on the recent email discussion on scheduling, my following comments.
 
First, MAC-protocols on rings are required to coordinate access to the shared medium in a fair manner. If one omits this protocol, one easily can construct cases where some of the nodes are unable to transmit. Thus, the MAC will be a crucial part of the standard.
 
For ring-speeds of 1 Gbit/s and above, the shared media can only exploited by simultaneaous access to the medium by many nodes and by destination removal).
Here, there are only two mechanisms, either a slotted transmission structure (ATM) or buffer insertion to cope with variable packet lengths (IP). In case of a slotted structure, packets have to be segmented at the transmit nodes and reassembled at the receiver nodes. In a ring with n nodes, each receiver node should have n reassembly machines. With insertion buffers (transit buffers) complete IP packet can be handled, and this is today the preferred way. Any thoughts in direction of  802.5 and FDDI for medium access make no sense.
 
The insertion buffers are in the transmission path and therefore all decisions have to operate at full link speeds. Buffers and complicated scheduling as in Diffserv at 10 Gbit/s, or in future even above that bit rate, are doable but require a lot of silicon space. Thus, are costly. It should be the goal to create a standard that allows a small insertion buffer size and rather simple scheduling. Three classes of traffic on the ring, as already discussed in emails, would fit completely. This already means that three parallel insertion buffers must be handled at link speed. This parallel structure is required to allow for overtaking of packets between the three priority classes. The jitter of the lowest class becomes of course larger, but the delay bounds of the higher class traffics can be guaranteed, which is the most crucial factor in IP-telephony and IP-multimedia. Other QoS policies should be done above the MAC. Ring nodes should be considered as nodes that have shared transmission links to all other nodes. Once the packet is given to the MAC or clocked onto the transmission medium, there is no need for level-3, etc switching control.
 
Having a long history in the design and implementaion of multi-Gbit/s shared-medium rings, the buffers in the transit path operate for me, without any discussion, always in cut-through mode. Once on the medium, the packet should be delivered to the destination with minimal interference by the access traffic of the same class. I recently learned however that in contrast to my assumption, the Cisco SRP operates their transit buffers as store-and-foreward. Each packet is completely received in the transit buffer before forwarding. Performance will show the difference in end-to-end delay. I like to point out that end-to-end delay will become the most dominant performance measure in the very near future.The requirements of the carriers will change when IP-telephony and IP-MM becomes a commodity and carrier-grade service have to be guaranteed in a simular fashion as in circuit-switching. When considering only 2.5 Gbit/s and 10 Gbit/s SONET/SDH or 10 GbE the difference will not be very crucial. However, for rings in the lower bit-rate region in public access, campus, or in-building areas, providing a huge market, the difference might have an impact on a broad success of the standard. Scheduling between the insertion buffers and the transmit buffers can in the same way be done both in cut-through and store-and-foreward mode. In principle, rings with cut-through and store-and-forward nodes having the same buffer structure would work together. Even scheduling between buffers could be different. What has to be the same is the MAC fairness mechanism.
 
Best  regards
Harmen R. van As
------------------------------------------------------------------
Prof.Dr. Harmen R. van As       Institute of Communication Networks
Head of Institute                      Vienna University of Technology
Tel  +43-1-58801-38800           Favoritenstrasse 9/388
Fax  +43-1-58801-38898          A-1040 Vienna, Austria
http://www.ikn.tuwien.ac.at      email: Harmen.R.van-As@xxxxxxxxxxxx
------------------------------------------------------------------