Feedback on bmwg-lanswitch-03 draft.
Kevin Dubray <kdubray@baynetworks.com> Wed, 26 February 1997 12:57 UTC
Received: from cnri by ietf.org id aa10474; 26 Feb 97 7:57 EST
Received: from newdev.harvard.edu by CNRI.Reston.VA.US id aa09330; 26 Feb 97 7:57 EST
Received: from netop3.harvard.edu (netop3.harvard.edu [128.103.205.103]) by newdev.harvard.edu (8.7.6/8.6.10-MT4.00) with SMTP id HAA07101 for <bmwg@ndtl.harvard.edu>; Wed, 26 Feb 1997 07:50:00 -0500 (EST)
Received: from mailhost2.BayNetworks.COM (ns1.BayNetworks.COM [134.177.3.20]) by netop3.harvard.edu (8.6.12/8.6.12) with ESMTP id HAA14003 for <bmwg@harvard.edu>; Wed, 26 Feb 1997 07:51:38 -0500
Received: from ns1.BayNetworks.COM ([134.177.1.107]) by mailhost2.BayNetworks.COM (8.8.5/BNET-96/12/06-E) with ESMTP id EAA29821
Received: from pobox.engeast.BayNetworks.COM (pobox.corpeast.baynetworks.com [192.32.151.199]) by ns1.BayNetworks.COM (8.8.3/BNET-97/01/07-I) with ESMTP id EAA23663
Posted-Date: Wed, 26 Feb 1997 04:50:55 -0800 (PST)
Received: from pestilence.engeast (pestilence [192.32.180.194]) by pobox.engeast.BayNetworks.COM (SMI-8.6/BNET-SVR4-S) with SMTP id HAA06865; Wed, 26 Feb 1997 07:50:56 -0500 for
Date: Wed, 26 Feb 1997 07:50:56 -0500
From: Kevin Dubray <kdubray@baynetworks.com>
Message-Id: <199702261250.HAA06865@pobox.engeast.BayNetworks.COM>
To: bob.mandeville@eunet.fr, bmwg@harvard.edu
Subject: Feedback on bmwg-lanswitch-03 draft.
Bob, It appears that your draft has undergone quite an evolution over the last two revisions. Here is my input on your latest version, draft-ietf-bmwg-lanswitch-03.txt. 2.4 Meshed traffic In my mind, it looks as you defined the concept of "meshed traffic" to mean "fully" meshed or what you originally proposed in earlier drafts as "multidirectional." And this appears consistent with the minutes of the last session. In retrospect, this may be an unduly restrictive definition. Indeed, section 16 of RFC 1944 (i.e., Multi-port) refers to a meshed distribution of traffic, albeit not a fully meshed distribution. (Scott, correct me if I'm wrong, but isn't the addressing suggested in RFC 1944, sec. 16 was what you were referring to when trying to offered the term "meshed?") Regardless, both "partial" and "full" mesh traffic paradigms are important, valid traffic distribution patterns in testing. So perhaps a base definition, followed by appropriate extensions are in order. For instance: 2.4 Meshed traffic: A traffic distribution pattern whereby multiple frames offered to an ingress port of a System Under Test are destined to two or more ports of that system. Discussion: In other words, frames offered to a port of a SUT in a meshed traffic distribution are not all destined to the same egress port of that SUT. 2.4.1 Partially meshed traffic: A type of meshed traffic distribution whereby multiple frames offered to an ingress port of a SUT are destined to a set of egress ports of that SUT. 2.4.2 Fully meshed traffic: A type of meshed traffic distribution whereby multiple frames offered to an ingress port of a SUT are destined to all other tested ports of the SUT with the exception of that ingress port itself. 2.x Non-meshed traffic: A traffic distribution whereby frames offered on a ingress port of a SUT are destined for a single egress port of the SUT in a one-to-one mapping. I think the above discussion is "backwards compatible" with historical references to "mesh" and "multi-port" but fully captures what you are trying to communicate with "multidirectional." As opposed to its use as a absolute benchmark, the power of the employing the "full-meshed" or "multidirectional" traffic model is its diagnostic capability. The use of the full meshed traffic model can suggest that additionally inspection of the SUT may be warranted. This idea has been offered at several BMWG sessions, but it has failed to be preserved in the current discussion. ================================================================== 2.9 Offered Load The definition of offered load says "..the number of frames per second that an external source attempts to transmit..." Later in the discussion, you suggest a more rigorous criterion by stating, "Frames which are not successfully transmitted by an external source ... MUST NOT be counted as as transmitted frames..." In my mind, your last statement contradicts the formal definition. If I set the transmit knob of the test tool to 14,880pps and the actual transmission rate is 14,773pps, which value do I report for offered load? The tool _attempted_ to transmit at 14,880, but could, for whatever reason, muster only 14,773. I would offer the suggestion to modify the wording of "attempts to transmit or address" to "successfully transmits." It is very important to absolutely describe the external stimuli presented to the SUT for action. You may wish to consult RFC 1242 and 1944 in light of your last paragraph in this section. The implication of suppressing the propagation of BPDUs or other like mechanisms may be counter to these documents, if my memory serves me correctly. 2.8 Port Load I think I know what you are trying to convey here, (i.e., define load with respect to port I/O), but I'm not so sure that I agree with your definition. The sticking point for me is where you say the port load is the number of fps that a DUT's port can be "observed to transmit and receive in response to an offered load." Ascertaining a tested port's ability to transmit frames is usually straightforward enough. I'm guessing that you're assuming the correct forwarding of a frame implies its successful reception by the test tool. No argument there. Without intrusive methods - that is directly relying on the DUT's statistics - determining whether a port "received" is not so straightforward. Consider the case where a frame offered to a DUT is actually read off the medium by the DUT's port but dropped internally by the DUT (e.g. clipped across the "switch fabric"). What signifies that the frame was correctly "received"? I think what you are really striving to communicate is the observed utilization of the medium attached to the port while taring any non-test specific traffic components from the measurement. 2.10 Balanced Load Again, I understand the concept because I am cognizant of the test condition that you are trying to describe. I've distributed the definition around to those both familiar with network benchmarking and those who are not. All suggest this definition should be made more clear. With regards to this discussion on types of loads, I noticed you dropped the concept of intended load that was offered in version 02 of your draft. While "intended load" doesn't appear crucial in your discussion of switches per se, I maintain (and the last BMWG minutes reflect) that it is important to differentiate the load the test technician asks the test tool to deliver from the observed load the test tool actually offers. The definition of intended load may coax the test technician to consider the following heuristic: if (offered load NOT_EQUAL intended load) then understand_why; Whether the reason why offered load differs from intended load is collisions, flow control, or the fact that perhaps the test tool just doesn't deliver the intended load, it is a important data point in the interpretation of the results. Furthermore, it's a datapoint that not all test technicians may consider scrutinizing up front. I would strongly advocate reintroducing the concept of intended load back into your draft. 2.12 Overload. In the definition section, you state overload as the condition where port(s) are loaded "in excess of the maximum rate of transmission allowed by the medium." Yet in the discussion section, you state "Port overloading... requires a minimum of two input ports and one output ports..." Doesn't the case where a test tool presents a DUT port with sub-minimal (illegal) IFG-spaced frames (in the case of Ethernet) fulfill your definition's requirement for overload? RFC 1242 presents overloaded behavior (section 3.12) as a systemic DUT condition; your definition presents a physical medium context. I think your definition may need to be reworked to reflect the spirit of your discussion. Or vice versa. ============================================================================= 2.13 Forwarding Rate In the presented discussion, use of the term "throughput" may be ambiguous in light of the definition of throughput as stated in section 3.17 of RFC 1242. Perhaps replacing "throughput" with "frame forwarding" may be more consistent? 2.14 Forwarding Rate at Maximum Load. It is not that uncommon to find test tools that CANNOT generate Maximum Load, as you previously define. Moreover, "Maximum Load" may vary with respect to other parameters other than a medium's theoretical, maximum utilization. For example, those media employing tokens, "maximum load" may vary as a function of Token Rotation Time, Token Holding Time, or the ability to chain multiple frames to a single token, etc. Recall that the last session of the BMWG suggested that the term "Forwarding Rate at Maximum Offered Load" (FRMOL) be defined. In the case where the max theoretical utilization of the medium can be generated by the test tool, max offered load IS maximum load (as you define). In the case where "maximum load" cannot be generated by the test tool, maximum offered load represents the upper (max) test bound. My counsel would be to reflect the BMWG's suggestion. The group also suggested defining "Maximum Forwarding Rate" (MFR) to be the largest forwarding rate from a set of forwarding rate measurements. It was further offered that comparison between MFR and FRMOL was useful in characterizing DUT behavior. ============================================================================= 2.16 Forward pressure While the current definition is adequate in describing a technique of violating Ethernet's backoff algorithm, perhaps it could be expanded slightly to reflect an increased scope of a network condition to be identified: Forward Pressure: A behavior whereby the SUT applies a heuristic that departs from or otherwise violates a defined, standardized protocol as a tradeoff for increased forwarding performance in those network conditions where congestion or contention is present. This definition would allow for the identification of the disregard of backoff algorithms, flow control primitives, or other medium capture constraints. Wording should be offered to cite that such behavior is unacceptable. ============================================================================= 2.18 Address handling Perhaps changing the term "address handling" to "address handling capacity" or something similar better conveys what it is that you are trying to describe? ============================================================================= 2.20 Flooding You seem to define a network condition here as opposed to a benchmark. If "flooding" isn't a benchmark, are the "Measurement units" applicable and/or necessary? 2.21 Illegal frames. Again this appears to be a term where you define a specific condition, that is, when frames are to be considered poorly formed with respect to the PHY/MAC layer. Being that the term resembles more a definition than benchmark, does the specification of measurement units make sense here? ============================================================================ Miscellaneous. Section 3, the Index of Definitions, is wonderful. However, it would have been my personal preference to present it earlier to the reader. Best Regards, Kevin
- Feedback on bmwg-lanswitch-03 draft. Kevin Dubray