A Tutorial on ccNUMA Using The Scalable Coherent Interface (SCI)
Tuesday, August 24, 1999, 1:00 p.m. to 6:30 p.m., Santa Clara University
IMPORTANT REMINDER - Early Registration deadline is Thursday August 19th.
The most interesting and important aspects of the SCI specification are
documented in C-code. Although most people retrospectively find that this
code is exceptionally comprehensive, it is time-consuming and difficult to
get through it alone. On August 24, SCIzzL will again offer an SCI
C-code tutorial which will uncover some of the obscurities; the session
will be presented by - B. Mitchell Loebel, President of MultiNode
Microsystems Company and Director of The PARALLEL Processing Connection.
Whatever are your objectives with SCI, this offering will definitely
shorten your "time to market" by several man years.
Who should attend? The agenda of the upcoming tutorial is specifically
tuned to assist those attendees who are planning to implement cache
coherence between compute nodes using the SCI protocol in the near future;
this will be a very practical session. And members of the instructing team
will be available to provide support after the seminar concludes.
Some of you have asked about the possibility of an all day seminar at your
site. That is definitely available. As for agenda, it can be tailored to
your needs; of course, we would need input from you as to how far along you
have progressed and what you are attempting to do. We think that attendance
should be limited to about 10-15 people, but we're flexible on that point.
We'll certainly cover more material, more deeply and interactively than we
will be able to do at the SCIzzL meeting. As regards background, it is safe
to say that there is very little of the code that we haven't explored over
the past two years and I am grateful for the able assistance of
Dave Gustavson and a number of exceptional people who are members in
The PARALLEL Processing Connection. Beyond merely sorting through the
C-code, our team has fully abstracted and documented the state machine
for the "Full Set" SCI protocol. The exact cost will depend on your needs
- please inquire at ++1-408-732-9869.
Details of the SCIzzL C-code seminar are as follows:
1. A familiarity with the C language is assumed.
2. A familiarity with the SCI specification (IEEE 1596) is assumed.
If you have a copy, please bring it with you. If you need to
purchase a copy, please contact IEEE standards sales
immediately, at 1-800-678-4333 (US) or ++1-732-562-3800
(fax ++1-732-981-9667); Or, through the Computer Society, at
1-800-272-6657 (outside California), ++1-714-821-8380 (inside
California or international), fax ++1-714-821-4641.
3. Completion of Dave Gustavson's Introductory tutorial or equivalent
experience is recommended but not essential.
THE TUTORIAL WILL EXTENSIVELY COVER THE SCI STATE MACHINE (WITH RELEVANT
STATE DIAGRAMS) IN ADDITION TO THE FOLLOWING MATERIAL. IN THE PAST, THE
INTENT WAS MAINLY TO GUIDE ATTENDEES IN GETTING THROUGH THE CODE. THIS
TIME AROUND, WE WILL SPEND MUCH MORE TIME DISCUSSING THE SCI STATE MACHINE
AND ITS INTERFACE TO SMP BUSES. THIS WILL BE THE BEST ONE WE'VE EVER DONE.
1. Simulation architecture -
The SCI spec is written in C and is encapsulated in a comprehensive
multithreaded simulation. In order to examine and understand the spec,
it is essential to understand the overall structure of the simulation.
2. Data structures -
Here we will describe the various data structures which are allocated
when the simulation starts up.
3. Packet movement between "chips" -
As an SCI packet moves the physical system which is described by
the simulation, threads are started, suspended and restarted.
Certain very important insights will be presented to help the
attendee get through the code.
4. Glossary of terms and functions -
The SCI spec is brilliantly done and the code is good. However,
the naming of functions and variables often leaves much to be
desired and is downright misleading. This seminar will definitely
shorten an attendee's path to overall understanding by defining
clearly the functionality of key routines.
5. "Double action" cache state machine -
SCI cache tag updates do not happen atomically, ie, a cache state
is set to an intermediate value when a packet is sent out and is
set to its final value when the response to that packet is
received. We will describe this important behavior in detail.
6. Conflict between concurrent transactions -
A. Within a node
In the case of several CPU's in a node (which seems to be the
way current system designs are going), certain actions by _some_
of the CPU's must be blocked. We will describe what blocking is
necessary and _why_.
B. External to a node
Similar to the first part of this section, but this will deal
with _inter_ node transaction conflicts (multiple processors in
a node) and interoperability considerations between dissimilar
7. Optimizing the SCI protocol within the specification.
8. Demonstration and interactive use of the SCI simulator.
Date: August 24, 1999
Time: 1:00PM to 6:30PM
Location: Santa Clara University, Santa Clara, CA - room 325.
Cost: $450 for members of SCIzzL who register by Thursday, August 19, 1999.
$500 per person for all others
Please make your check payable to SCIzzL and mail it to:
The PARALLEL Processing Connection.
1066 Sunnyvale-Saratoga Road, Suite 18
Sunnyvale, CA 94087, USA
fax (for VISA or Mastercard): 408 732-9869
(Please email Dave Gustavson - firstname.lastname@example.org to inquire about
SCIzzL membership.) On-site registrations and those received after
August 19, 1999 will be at the undiscounted price. Additionally,
handout materials can only be guaranteed for pre-registrants; others
may have to be mailed. For those who need additional time to move
paperwork within their organizations, we will take a "letter of
intent to register" from the organization received on or before
August 19th as an indication of pre-registration. All registrations
must be paid in full in advance of attendance.
online registration form
or email the following info to email@example.com:
Registration form -
Snail mail address:
B. Mitchell Loebel
CEO, Chief Technical Officer
MultiNode Microsystems Corporation ++1-408-732-9869
The PARALLEL Processing Connection ++1-408-732-9869