LCF_ad_hoc

Dal Allan dal at endl.com
Fri Nov 4 16:51:46 PST 1994


              Minutes of the November 2 Low Cost Fabric Ad Hoc 


The Low Cost Fabric Ad Hoc authorized by X3T11 which met at the Denver 
Stapleton Red Lion on November 2 was attended by the following:

 Dal Allan       ENDL                    Jean Kodama     QLogic 
 Charles Binford AT&T GIS/NCR            John Lohmeyer   AT&T GIS/NCR 
 Kurt Chan       Hewlett Packard         Kumar Malavalli Hewlett Packard CNO 
 Jim Coomes      Seagate                 Bob Mayer       Hewlett Packard 
 Richard Dugan   Hewlett Packard         Jim McGrath     Quantum 
 Dave Ford       Cambex                  Robin Purohit   Hewlett Packard CNO 
 Giles Frazier   IBM                     Bob Snively     Sun Microsystems 
 Gene Freeman    AT&T GIS/NCR            Ken Thompson    AT&T GIS/NCR 
 Stillman Gates  Adaptec                 Horst Truestedt IBM 
 Norm Harris     Adaptec                 Fanny Wong      IBM 

The minutes of the ad hoc are being distributed on the SCSI and Fibre 
Channel reflectors, as the attendees were a mix of both groups. 

The objective of the ad hoc was to table all the ideas which are presently 
circulating on how to configure Fibre Channel for the best performance at 
the lowest cost. Once a base which included all the ideas was established, 
individuals would be volunteered to pursue those in which they had an 
interest, and report on their merits. 


Heterogeneous Operation:  Jim McGrath was the only attendee present who had 
come prepared with a presentation. A joint press release of Quantum, Seagate 
and Hewlett Packard was included in the handout, in which the three 
committed to Fast20 and Fibre Channel Arbitrated Loop for SCSI disk drives. 

The primary point Jim made was that to become a mainstream interface, Fibre 
Channel had to be able to connect devices of heterogeneous speeds. Although 
this capability exists with switches, it is not cheap enough. 

He emphasized that Fibre Channel had to be popularly priced, and not be seen 
as the preserve of supercomputing and high end workstations. At present, the 
FC-AL has no provision to share devices of various speeds, and is perceived 
as a 100 MBs solution only, and not at low cost 25 MBs. 

As the least expensice fabric alternative, Jim wanted FC-AL to provide the 
equivalent functionality to mixing SCSI-2 and FAST-20 devices on the same 
bus. He proposed alternatives such as small loops combined by a low cost 
switch, and a way to expand the FC-AL protocol to alternate between 25 MBs 
operation on the loop and faster speeds between faster ports. 


Multiplexed Low Speed Ports:  Dal Allan presented hardware considerations 
and the application possiblities for an alternative to the loop which used 
radial connects to low speed devices and relied on buffers to speed match 
and multiplex transfers over to high speed ports connected on the host side. 


F_Ports:  Robin Purohit described the considerations involved in the design 
of an F_Port which was attempting to provide maximum performance between two 
loops on opposite sides of a switched fabric. The problems of buffering on 
the receive side and flow control were expensive in management and 
buffering. 

Horst Truestedt described the opposite extreme in which loop devices could 
transfer only one frame every arbitration cycle through the F_Port to the 
fabric. This would simulate the behavior of a single N_Port being serviced 
by an F_Port plus the overhead of arbitration between frames. This reduced 
the complexity of the F_Port design but at an unknown performance loss. 


10-bit Interface:  The possibility of using 10-bit interface between devices 
physically close to each other was described by Kurt Chan and Dal Allan. 
Kurt felt it might be possible to use the 10-bit interface currently being 
designed but Dal thought the bus considerations required a layer which 
accepted the 10-bit interface and adapted it to bus operation. 

The immediate objective would be to achieve Gbaud operation at 1/4 speed 
costs by eliminating the expense of serial transceivers. As the cost of 
serial components fell and Gbaud became integrated, the parallel interface 
would advance to higher transfer rates. 


Summary:  All of the above were actively debated by all attendees, and there 
were contrary opinions expressed on every proposal. It was a very active and 
highly participative morning. It cannot be said that any conclusions were 
drawn or agreed to but we had a lot of fun.... 


Cost Elements:  After lunch, cost bogies were established for mid-1996 
availability of 25 MBs copper interface elements assuming Q100,000 ordering 
by disk drive vendors of their parts and other units scaled accordingly. 

   +---------+       +---------+       +---------+       +---+-----+
   |   Host  |       |         |       |Packaging|       |FCP|Disk |
   | Adapter |       |  F_Port |       |  F_Port |       |   |Drive|
   +---------+       +---------+       +---------+       +---+-----+
    $70 - $150          $150                ?             $10   ?

Jim insisted that an FC-AL front end operating at 25 MBs had to be priced 
the same as single ended SCSI, and that was only $10 of the drive cost. Jim 
was assuming in-house cost of an FC_AL front end (to avoid paying merchant 
vendors the margin they needed to justify building a component). 

There was a wide divergence of opinions on how cheaply an adapter could be 
built with Bob Snively at $70 and Norm Harris at $150. 

Kumar Malavalli and Robin Purohit felt that to be widely adopted, an F_Port 
had to be built at a cost point of around $150. 

The price of these elements to OEMs were expected to be 3-4 times as high, 
depending on the amount of R&D recovery. 

No quantitative analysis is involved in these numbers. The purpose of the 
exercise was to instill consciousness, awareness and quantify 'low-cost'. 
These numbers were chosen by arguing amongst each other until the hubbub 
subsided to an acceptable level. 


Mixed Speed Solution:  Jim McGrath ended the day by leading a design effort 
to modify FC-AL to intermix devices of multiple speeds. His ideas ran into 
strong opposition as severely impacting the present protocol and assumed 
topology. 

The alternative with minimum time to market and minimum impact on standards 
was to operate a dual ported initiator running one loop at low speed and the 
other at high speed. If redundancy was needed, use dual ported devices and 
connect them to another dual-ported initiator. 



Action Items: 

Robin Purohit - assess cost impact and performance achievable by a single 
frame transfer every arbitration cycle through an F_Port to a host. 

Robin Purohit - assess cost and performance achievable by an F_Port designed 
to provide up to 80% of loop bandwidth on one side of the fabric to a host 
on the other. 

Horst Truested/Jim Coomes - assess performance impact of switching speeds 
based on initiator control of the loop speed, using the LR circuit as a way 
to bypass low speed devices during high speed transfers, and vice-versa.

Kurt Chan - estimate cost considerations of packaging and components for a 
10-bit bus version of FC-AL. 

Norm Harris/Stillman Gates - estimate considerations/costs involved in a 
multiplexing adapter which has radial connects to devices. 



Next Meeting:  None scheduled. 

The above action items and further LCF discussions will be part of the 
Fibre Channel working group, scheduled for Thurday of X3T11 plenary week. 

If other ad hocs are presumed necessary to pursue specific ideas/projects, 
they will be scheduled by someone who wants to lead such an activity. 

The time for completion of the above action items and report on them is 
December 8 at the San Jose Red Lion. 



My thanks to all who attended. Please advise of any errors or omissions so 
they can be corrected before these minutes are submitted in my reports to 
X3T10 and X3T11. 


I. Dal Allan





More information about the T10 mailing list