FC-AL Direct Attach Disk Minutes, 12/94

Kurt Chan kc at core.rose.hp.com
Wed Dec 14 14:05:43 PST 1994

           FC-AL Direct Attach DIsk Ad Hoc Meeting Minutes
                      Red Lion San Jose

  Dal Allan, ENDL                     250-1752 at mcimail.com
  David Allen, TI                     daln at ti.msg.csc.ti.com
  Jean Baudrexl, Intellistor/Fujitsu  jbaudrexl at intellistor.com
  Charles Binford, ATT-GIS/NCR        charles.binford at wichitaks.ncr.com
  Kurt Chan, HP                       kc at core.rose.hp.com
  Vince Cavanna, HP                   vvc at core.rose.hp.com
  Mehmet Cirit, Adaptec               mcirit at eng.adaptec.com
  Jim Coomes, Seagate                 jim_coomes at notes.seagate.com
  Gene Freeman, ATT-GIS-NCR Micro     gene.freeman at colospgs.ncr.com
  Mike Fry, NCR Micro                 mike.fry at ftcollins.ncr.com
  Ed Frymoyter, HP FCSI               70523.3010 at compuserve com 
  Ed Gardner, Quantum                 gardner at acm.org
  Stillman Gates, Adaptec             stillman at eng.adaptec.com
  Bill Gintz, Conner                  bill.gintz at conner.com
  Norm Harris, Adaptec                nharris at eng.adaptec.com
  Larry Lamers, Adaptec               ljlamers at aol.com
  Hyeun Tae Lee, ETRI Korea           htlee at prism.etri.kr
  Allison Parsons, Conner             allison.parson at conner.com
  Craig Rich, Cypress                 csr at cypress.com
  Colin Schaffer, Intellistor/Fujitsu cschaffer at intellistor.com
  Marc Shoquist, Computing Devices    marc.c.shoquist at cdev.com
  Bob Snively, Sun Micro              bob.snively at sun.com
  Craig Stuber, Adaptec               stuber at eng.adaptec.com
  Sassan Teymouri, Adaptec            sassan at eng.adaptec.com
  Lloyd Thorsbakken, Unisys           let2 at unisys.rsvl.com
  Horst Truestedt, IBM                truested at vnet.ibm.com
  Peter Walford, Demografx            walford at btr.com

Craig Stuber from Adaptec shared some performance numbers from simulations
of PCI host adapters connected to FC-AL (both single loops and dual loops).

PCI configuration:
- 64 bits at 33MHz
- 128 byte cache line
- 7 clock arbitration time.
- No request/grant delays or data transfer wait states in single-loop config
- Dual loops had contention for PCI bus

Disk/link configuration:
- 0.5m cables
- 8ms access time
- Command queueing not invoked (one command at a time per drive)
- 200us drive context switchtime

- Peak performance achieved with 55, 35, and 25 drives for Full, half, and
  quarter speeds.
- Single loop FC-AL (50 drives) running at Gb can sustain 73MB max at 64k
- Frame size has little effect on throughput
- Single loop (50 drives) running at quarter speed can sustain 22MB/s
  with 64k byte blocks
- 64-bit, 33MHz PCI can accomodate 3 single-loop FC-AL Gbit controllers
  before PCI becomes the bottleneck. Rumors - 66MHz won't be 2x performance
  due to system memory limitations (memories must be run at 64 bit widths
  and a bus rate). 
- Dual loop (25 drives per) at quarter speed can sustain 42 MB/s at 64k
- Splitting same number of drives into 2 loops at Gb delivers 25% more 
  MB/s and 10% more IOs/sec
- Splitting same number of drives into 2 loops at quarter speed delivers
  30% more MB/s and 40% more IOs/sec.

Observation: optimum number of drives will lower as tagged queueing added
(more I/O's per drive). More cost effective - fewer spindles to do same
benchmark, and since fewer nodes latency will be slightly less.

For the purposes of determining E_D_TOV, private loop devices should
consider themselves in a "point-point" configuration.  That is,
E_D_TOV is discovered through PLOGI, not FLOGI or RTV (and R_A_TOV is
defined as 2 x E_D_TOV on a point-point link).

E_D_TOV by itself is not used anywhere in the document.  The optional
Initiator E_D_TOV timers referenced in section 11 were removed.

Vince Cavanna made his case for coax as the cable of choice for bulkhead
to bulkhead connections. Some assertions:

- Hirose still has the possibility of providing GLM compatibility.  But
  it is still single-sourced and not as pervasive as BNC.  
- Coax has lowest cost structure.  
- Physical loop structure desireable (DB-9 cannot - only supports pt-pt).  
- Coax was designed for RF - so we're not stressing the cable.  
- Coax has several choices of cables/connectors and adapters (taps,
  50-75 ohms, etc).  
- Video applications are using coax to 1.5Gbaud, and use a segment of
  coax as a common mode choke.  As with 8b/10b, there is no DC
  component in the video signal.  A common mode choke isolates chassis
  ground from shield.
- Coax and SMA connectors compatible with no signal loss, they're also
  cheap.  SMC also comes in 75 ohms.  There doesn't appear to be a
  significant difference in performance between 50 and 75 ohms.
- Power budget not sufficient to drive long runs of bypassed ports.
  Receiver is always TNC, transmitter is always BNC to avoid
  mis-polarization.  SMA cannot do this.  
- Twinax should have better EMI than coax (which is not perfectly
  shielded?).  Connection between cable shield and connector housing
  is not 360 degree in twinax.

There was far from unanimous agreement on all of these points, and
therefore another meeting will be needed to continue this discussion.

The next meeting will be at Adaptec in Milpitas on Thurs, Feb 2 at
830am.  The meeting will be almost exclusively devoted to continuing
the discussion of FC-0 issues.  No other agenda items are planned so
far.  There's a Crown Sterling Suites in Milpitas at 901 Calaveras
Blvd, 408-942-0400. It's 7 miles from the SJ airport, with complementary
shuttle service. Contact Norm Harris for directions to Adaptec:
nharris at eng.adaptec.com.

I have proposed that cable/connector standardization be continued in
the Small Form Factor (SFF) committee which meets during X3T10 weeks
and specializes in the standardization of physical plants for storage
devices and subsystems.  The first SFF meeting after Feb 2 will be
Fri, Mar 10 in Newport Beach, CA (Orange County).  At that meeting,
SFF will assume the charter of standardizing FC copper bulkhead
implementations.  Contact Dal for details on SFF.

The disk profile will reference the appropriate SFF document.  Version
1.2 postscript will be available on ncrinfo by Fri, 12/16.

Since no consensus was achieved whether ACA behavior should be Required
vs Allowed in compliant Targets, the profile shall only state the following
regarding ACA:

- During the discovery process, Initiators shall use INQUIRY to discover
  whether or not Targets support ACA behavior.

- NACA=1 shall be set in the CDB control byte only when the sending
  Initiator and receiving Target both support ACA functionality

- NACA=0 shall be used when either the sending Initiator or receiving
  Target does not support ACA behavior.

Future work with X3T10 on simplifying/clarifying ACA behavior for
disks will be left to individual companies.  I am receptive to any
comments in this area, and if an attractive proposal comes up we can
continue the discussion via the disk_attach reflector.

More information about the T10 mailing list