| This is a general distribution memo, so I'm posting it.
From: NAME: Susan Kenney
FUNC: NPB Sales Communications
TEL: 226-5081 <KENNEY.SUSAN@A1GLRMAI@TAY>
To: See Below (Distribution List Truncated)
TO: WORLDWIDE NPB FIELD ORGANIZATION
CC: NPB INTEREST LIST
*******************************************************************************
THIS NPB SALES FLASH IS FROM STAN KOPEC, NPB MARKETING,
HUBS/COMPETITIVE ANALYSIS. IF YOU HAVE QUESTIONS, PLEASE CONTACT
STAN AT DELNI::KOPEC OR DTN 226-6404.
THIS MESSAGE IS 9 PRINTED PAGES.
*******************************************************************************
***Digital Confidential**
The following represents a Digital-NPB evaluation of the February 1996
Data Communications (DC) switching article entitled "Canned Heat". This memo
is Digital Confidential and is intended primarily for Digital NPB employees.
The DC test covered the following main areas of evaluation using either Fast
Ethernet or FDDI as the high speed backbone or switch interconnect:
Frame Loss and Per-port Throughput under extremely heavy loads, Head of line
blocking, Error Handling, and Latency over the backbone. The testing was
done at the European Network Laboratory (ENL) by Robert Mandeville.
David Newman was the test editor.
Introduction (details in main document)
=======================================
Digital's comparative results were very positive for Per-port Throughput,
Head of line blocking and Error Handling. Frame loss and Latency results from
this test were not as positive and require an in-depth discussion of how these
switches were tested. There were also several inaccuracies in the article
which affect the customer's perception of Digital's switching products.
This memo will:
o Correct the stated inaccuracies
o Expand the printed discussion of Fast Ethernet vs FDDI in the
backbone to a more complete discussion including all of
the issues that Network Managers should be aware of
o Comment on particular statements in the article concerning both
Digital and our competitors' products
The main issues which will be addressed include:
Issue: DC statement that GIGAswitch/FDDI had latency up to 180 microseconds.
Sales message: This conclusion is absolutely false. DC incorrectly interpreted
the results from their test. GIGAswitch/FDDI's latency numbers are on
the order of 15 microseconds for switched FDDI packets which is the lowest
latency among FDDI switches on the market today!
Issue: DC claim that Fast Ethernet is the preferred backbone technology vs FDDI.
Sales message: Fast Ethernet is certainly a viable backbone technology.
However, Network Managers need to look beyond single-test comparisons such as
latency (which will be lower than FDDI due to no translation) and consider
the Reliability, Availability, Standards Compliance, Maturity, Scalability,
Interoperability and Flexibility of FDDI networks as well.
Issue: DC acknowledgement that more aggressive switch Ethernet controllers
will result in improved test results.
Sales message: While more aggressive switch Ethernet controllers may result
in considerably improved test results, the article acknowledges that this
would not be a feasible solution in a "heavily loaded production network".
(page 86 of article)
Issue: Many switches may attain very low frame loss by using backpressure,
a congestion control mechanism.
Sales message: Backpressure may be used to minimize frame loss however the
article acknowledges that "There is compelling evidence that backpressure can
actually hinder performance." (page 92-94 of article). This can adversely
affect not only switch throughput, but also attached shared LANs.
Issue: How could overloads occur in a real world network?
Sales message: It is certainly the case that switches in this test were
severely overloaded in the case results which were published. However, it is
not possible to overload a real-world network or switch to the extent depicted
by this test (150% offered load, fully-meshed, 64 byte only, burst size of up
to 744 packets) by any example sited in this article. Furthermore, there
has been no documentation of such an environment outside of a test laboratory.
GIGAswitch/FDDI latency
=======================
The first issue which must be corrected immediately is the DC statement
(page 96) that "DEC's GIGAswitch, with per-frame latency of nearly 180
microseconds...". This is absolutely false. DC incorrectly interpreted
the results from their test. The GIGAswitch in cut-through mode yields on the
order of 15 microseconds latency for FDDI switched packets. The DC test method
of measuring this latency is FIFO. This 15 microsecond latency has been
measured in Digital's lab, and others, and we will be discussing this with Data
Communications as soon as possible.
Stephen Saunders, another Data Communications editor, stated in his November
1995 Data Communications article entitled "FDDI Switches: Immediate Relief
for Backbones Under Pressure"...
"When the Gigaswitch/FDDI is running in cut-through mode, the port that
receives a frame transmits it to the destination port as soon as the first
60 bytes of the frame have been received--limiting delay to 15 microseconds,
according to DEC. Only one other FDDI switch comes close to matching the
Gigaswitch/FDDI in cut-through capabilities: NSC's Enterprise Routing Switch
can be set to start forwarding frames after the first 200 bytes have been
received, keeping latency to between 50 and 60 microseconds, regardless of
frame size."
Another quote from the same article reinforces Digital's customers' acceptance
and realization of the benefit and effectiveness of the GIGAswitch/FDDI product
in the most demanding of environments. "With nine vendors selling FDDI
switching gear, there's no shortage of equipment to choose from. Digital
Equipment Corp. (DEC, Maynard, Mass.) is at the head of the pack; its
Gigaswitch/FDDI product accounts for 70 percent of all FDDI switches sold."
Finally, in the August 1, 1995 issue of Network Computing, Art Wittman
(after running a series of performance tests) states "...the GIGAswitch/FDDI
is truly a marvel and for busy FDDI-based networks, nothing else compares."
Sales Message: GIGAswitch/FDDI has the lowest latency among FDDI switches
on the marketplace today, and "nothing else compares" for busy FDDI
networks.
Fast Ethernet vs FDDI as a Backbone Solution
============================================
The tone of the "Canned Heat" article strongly pushes Fast Ethernet as the
preferred backbone technology vs FDDI "for high speed performance with minimum
frame loss...". While no one would argue with the statement that
Ethernet-to-Fast Ethernet switching requires "no conversion... of frames"
and thus will result in slightly lower latency than Ethernet-to-FDDI
switching, what is more important to most Network Managers is how a
backbone switch's Performance Scales as more LANs and users are
added to the switch -- see Throughput and Scalability discussion below.
On the issue of frame loss, this can occur for many reasons but using FDDI vs
Fast Ethernet as a backbone is NOT the primary factor. Both technologies
can handle very high traffic loads. There are many ways of controlling Frame
Loss, one of which is the Backpressure method which the authors themselves
state "...(there is) compelling evidence that backpressure can actually hinder
performance". It would be an appropriate question to ask of any customer if,
in real-world situations, it is more important to minimize frame loss and
receive lower throughput, or to have a minimal amount of packet loss and
receive higher throughput?"
What is typically more important to Network Managers is Throughput, which is
summarized on page 90. This graph places Digital's DECswitch 900EF firmly in
the same top tier (above 4000 pps) of per-port throughput performers as Cisco,
Crosscomm, and Madge. The second tier includes Bay, Fibronics and 3Com.
Knowledgeable Network Managers who are really comparing Fast Ethernet to FDDI
as a backbone technology seek out the complete story including the following
facts about FDDI.
FDDI is a mature, well-understood standard and a widely implemented backbone
technology that provides innate management (SMT) capabilities as well as:
-Built-in Fault-Tolerance (Dual Rings)
-The ability to very easily accomodate any building or campus's
physical characteristics (Ring and/or Tree topologies)
-Supports a greater extent (up to 100KM ring)
-Supports dual homing of stations which is an additional unique
capability providing increased Availability
Moving beyond the pure technology discussion, when a Network Manager implements
one (or several) GIGAswitch/FDDI's and DECswitch 900EF/PEswitch 900TX
switches, he/she is gaining the benefit of the following very important
attributes:
a) Scalability
-Each GIGAswitch/FDDI supports up to 34 ports linked with
patented crossbar switch technology
-Multiple GIGAswitch/FDDI's may be linked together to
provide a larger number of central switched ports
b) Modularity and Flexibility
-Place switches closer to your users/departments when run in
standalone/stackable mode or integrate the same switches into
a DEChub 900MS chassis in wiring closet or computer room
-Optionally upgrade the 900EF's to IP or Multiprotocol
distributed routing
-Use lower cost PEswitch 900TX for smaller workgroups
c) Availability/fault tolerance
-Innate characteristics of FDDI mentioned above
-Redundant power supply option both in-hub or out-of-hub
d) Robust switching
-Spanning tree supported on all switches
-8,000 addresses on 900EF, 16,000 on GIGAswitch/FDDI.
(Note that performance tests were NOT run with large
numbers of addresses - as advertised - on the switch.
This negates a distinct advantage that switches which were
designed to support large enterprises and departments have.)
e) Standards compliance
-IEEE 802.3, ANSI X3T9, SNMP, SMT
f) Unique features
-Priority queues for Spanning Tree and SNMP traffic on
DECswitch 900EF and PEswitch 900TX to ensure a stable and
manageable network
-Self healing rings and trees as well as automatic reconnection
into backplane FDDI ring upon reinsertion in the
DEChub 900MS backplane
-Full duplex FDDI supported on all Digital switches in test
providing greater bandwidth
-Rate limiting feature on DECswitch 900EF/PEswitch 900TX
providing control and management of broadcast storms
-ARP server capability of GIGAswitch/FDDI which prevents
broadcasts ARP requests from flooding the FDDI switched network
-Low cost SNMP/GUI management (HUBwatch for Windows)
which manages all devices
g) Near-future enhanced capabilities with DECswitch 900EF and
GIGAswitch/FDDI working together in synchronization as well as
future switches (see enVISN PID which will be coming out in
March 1996)
-Multiple Classes of VLAN membership
-Central policy based administration
-Increased security
-Distributed routing, multilayer switching
-Scaleable bandwidth
-Support for new services (e.g. multimedia)
-Integral RMON
As far as performance goes, feel free to reference the November 13, 1995
Communications Week article by Ed Mier and Robert Smithers Jr. Included is
an excerpt from a previous email summarizing the DECswitch 900EF's
winning performance in a real-world client-server test where
the FDDI backbone was loaded to a more realistic 90% of load as opposed to the
non-real-world 100-150% load under ENL's test. In Mier's performance
test, Digital's DECswitch 900EF and 3Com's LANplex 2500 tied for best
overall performance, and "Digital's DECswitch 900EF could consistently pass
91 percent (of traffic)" for maximum Ethernet packet sizes.
What follows (headed by !) is an excerpt from a November 95 memo
summarizing some of the important points in the article. Note that
unfortunately there was no overlap between the competitors in the two articles;
however, the benefits of Ethernet-to-FDDI switching and FDDI as a backbone
technology as well as the performance of the DECswitch 900EF are very
clearly stated.
Note: The Cisco Catalyst 2800 submitted to Mier was configured as an
Ethernet-to-FDDI switch while the one submitted to ENL was
Ethernet-to-Fast Ethernet.
!The DECswitch 900EF was submitted in head to head competition with 5 other
!switches: 3Com Lanplex 2500, Cisco/GJ Catalyst 2800, Plaintree Waveswitch 100,
!RAD FEB-4 and SysKonnect SK-NET Switch 6608.
!
!Test methodology concentrated on the following areas of evaluation:
! Installation/configuration
! Ease of use, UI, doc and on-screen help
! "Basic features"
! "Advanced features"
! Management and Administration
! Performance (response times & timed tests for transactions & file xfer)
!
!The 900EF came in "a close second place" overall to the 3Com box and there
!are a number of very positive comments throughout the article on the EF in
!particular. Cisco/GJ Catalyst 2800 came in 3rd.
!
!Some brief comments on the article...
!
!1) 3Com 2500 was "most expensive", "no auto purging of entries in address
! tables when it becomes full but rather floods all new addresses on all
! ports"(!)
!
!2) Cisco Catalyst 2800 "drops user data"(!) when the testers injected bad crc
! FDDI packets into the mix..., "Windows based management application is
! under development but was not ready for testing..."
!
!3) 900EF "has a rich set of filtering capabilities", "no problems handling
! errored packets", "EF could consistently pass 91% of Ethernet's maximum
! load with max sized packets",...
!
!4) The last line in the article stated "3Com's and Digital's switches stood
! just a little taller than the remainder."
Interoperability was also a key focus of the Mier article. This of
course is a direct result of the maturity of the technology. For those
risk-averse Network Managers, Mier/Smithers noted "...there were no
interoperability problems between any of the switch combinations."
Sales message: FDDI-to-FDDI and Ethernet-to-FDDI switching, and in
particular the GIGAswitch/FDDI and DECswitch 900 family of products, are a
solid choice for Network Managers interested in Reliability, Availability,
Standards Compliance, Maturity, Scalability, Interoperability, Flexibility and
Performance!
Issues to be aware of in the article
====================================
3Com LANplex 6004 (Page 86)
---------------------------
"Initially, the unit junked a lot of frames under bursty conditions. The
patch 'fixed' the problem by prohibiting other stations from transmitting when
the LANplex had data to send. That's OK for the testbed, but try it on a
heavily loaded network and the number of timeouts will skyrocket as attached
stations attempt to retransmit data."
Comment: The patch that 3Com supplied helped them receive excellent scores on
the performance charts however this raises the question of... Why
was a vendor allowed to affect the characteristics of their switch's Ethernet
controller to "become overly aggressive" so that they could do well on the
tests, even though it would not be a feasible solution in a "heavily loaded
production network"?
Sales message: The important point for the customer to realize is that
products that are designed or adjusted to pass test suites are not necessarily
the right choice for "heavily loaded production networks." Digital's
products are designed with the customer's real-life environment in mind, not
a theoretical worst case scenario that has never been observed outside of a
test lab. To see an example of a test methodology that attempts to emulate a
real-life environment, see the November 13, 1995 Communications Week article
entitled "Ethernet-to-FDDI Switches -- Linking LANs".
Backpressure (Page 92-94)
-------------------------
"(There is) compelling evidence that backpressure can actually hinder
performance."
Comment: 3Com LinkSwitch 1000 and Madge LANswitch implement backpressure as
a congestion control mechanism. Bottom line, although raising carrier or
creating an artificial collision on an attached shared LAN that wishes to
transmit will minimize packet loss THROUGH THE SWITCH as discussed above, it
will also adversely affect overall throughput through the switch and (perhaps
more importantly) adversely impact communication between stations on the same
shared LAN!
In fact, 3Com's LinkSwitch 1000 technical documentation states the following
warning: "Note: If you connect repeated segments to the LinkSwitch 1000 and
the stations on the repeated segments are running a peer-to-peer protocol (for
example Windows for WorkGroups), we recommend that you disable Intelligent
Flow Management on the ports to which the repeated segments are connected."
(Intelligent Flow Management is 3Com's Marketing name for Backpressure.)
Sales message: In real-world environments, isn't it more important to get
higher throughput through the switch with some minimal level of packet loss
and not affect the attached shared LANs vs using backpressure as a congestion
control mechanism?
Real-world Overloaded networks (Page 92)
----------------------------------------
"In real-world networks, overloads are prone to occur simultaneously on many
ports (such as when numerous users log in first thing in the morning)."
Comment: This is a very weak argument for trying to justify (as real-world)
the 100-150% offered load, bidirectional, fully-meshed, 20x20,
64-byte only, up-to-744-packet-bursts suites developed for this test. One
piece of information that should assist in helping your customer understand
that this testing environment is non-real-world, is the fact that no one has
observed or documented any real-world company, real-world application mix, or
real-world network configuration that would come close to simulating the
environment described by this testbed.
Sales Message: Although there are some aspects of the test methodology which
simulate real-world environments, of what value are the results from a
NON-real-world case described above? What exactly is being tested here? Is
it not more important to be able to handle an environment where packet sizes
and bursts are distributed across a realistic range (dependent on the users'
applications) as opposed to a worst case scenario that has never been
documented or observed in any other location than a test lab?
See November 13, 1995 Communications Week for a much more real-world testing
methodology.
Inaccuracies that need correcting within the article
====================================================
Page 82-83
a) Statement: Table stated that Cut-through answer is "no".
Correction:
GIGAswitch/FDDI does of course support cut-through switching.
It also supports store and forward as an option (settable by user).
DECswitch 900EF and PEswitch 900TX are store and forward devices.
b) Statement: PEswitch 900TX supports 5 Ethernet MAC addresses
Correction:
PEswitch 900TX supports a maximum of 64 Ethernet addresses today
and will support 96 Ethernet addresses in Q4 FY96.
c) Statement: Shared memory bus architecture
Correction:
GIGAswitch/FDDI has a patented crossbar switch matrix backplane.
DECswitch 900EF and PEswitch 900TX uses a shared memory approach.
d) Statement: $37,950 and $20,980 prices
Correction:
The list price as described by footnote #2 on page 82 is:
$33,380 for 4 DECswitch 900EF with 4 DEHUA-CA
$17,780 for 4 PEswitch 900TX with 4 DEHUA-CA
Note that it looks like DC took this cut at a "common" configuration
because they had so many different vendor configurations.
To Distribution List:
_nm%mkots1::azcona@A1GLRMAI@TAY,
_nm%chefs::collinsm@A1GLRMAI@TAY,
_nm%genie::dincau@A1GLRMAI@TAY,
_Nm%mkots3::friedland@A1GLRMAI@TAY,
_nm%delni::gayowski@A1GLRMAI@TAY,
_nm%cgooa::goldsmith@A1GLRMAI@TAY,
_nm%delni::jscott@A1GLRMAI@TAY,
_nm%stkhlm::klarsson@A1GLRMAI@TAY,
_nm%warder::purnellr@A1GLRMAI@TAY,
Distribution List Truncated
|