Software Systems
Spring 2005

Outline:

0) Homework 1

1) Patterson's article

2) the system model

3) example: disk drives



For next time:

1) look over the three Disk Drive handouts

   SCSI interface
   pages 680-686 from Hennessey and Patterson
   Cheetah product manual

   Keep these handy.

2) finish Homework 1

3) prepare for a quiz covering Homework 1


Homework 1
----------

What have we learned so far?

Questions?

Let's do Sections 1.3 and 1.5


Patterson's article
-------------------

Comments?  Questions?

I started the semester with this article

1) because it says interesting things about latency and bandwidth

2) because it introduces several of the themes we will be talking
   about all semester.

Of course, not all of it will make sense at this point.

1) bandwidth improves faster than latency

2) Moore's law

   fast transistors and more pins, good for bandwidth

   fast transistors, good for latency

   bigger chips, bad for latency

3) physical lower bounds

   speed of light in networks

   capacitance on and between chips

4) marketing

5) improving latency usually improves bandwidth

6) sometimes improving bandwidth hurts latency

   (Patterson may be a overstating his case here.  There are also
   many ways to improve latency that consume bandwidth -- he mentions
   them later.)

7) layered design hurts latency

   (operating system overhead)


Tools for improving latency:

1) caching

2) replication

3) prediction


Lessons for innovators:

1) if you have an idea that needs a little more bandwidth,
   you won't have long to wait.

2) if you need lower latency, the wait may be longer, or
   infinite.

3) if trends continue, desperate latency-reducing strategies
   may pay off



System Model
------------

Keep in mind, as always, that a model is a kind of lie.

The parts of the system that we care about (for this class) are:

1) CPU

2) registers and other hardware state

3) cache(s)

4) memory 

5) buses and ports

6) disk drives

7) network interfaces

8) graphics cards

9) human interface devices

10) measurement and control devices (A/D and D/A)


For each component, the questions we are interested in are:

1) what is the software interface?

2) what are the relevant performance characteristics?

3) is there a model for this component that is sufficient
   for our purposes as users/designers of software systems?

4) can we build an abstract interface for this device that will
   make it easy to use without sacrificing control and performance?


People who write device drivers need (1)

People who design and analyze systems need (2) and (3)

People who design operating systems and programming languages
work on (4)


The disk drive
--------------

We will answer the previous questions using three carefully-selected
readings.

The software interface to a SCSI disk drive is... um... SCSI.

SCSI (Small Computer Systems Interface) is one of several interfaces
cpus use to talk to disk drives.

As a quick aside, SCSI and many of the other interfaces in
Intel-based systems are OPEN STANDARDS, which means that many
companies can manufacture devices that will work together 
interchangeably.  As a result, these manufacturers are forced
to compete on the basis of performance and cost, the result of
which is an ENORMOUS benefit to consumers.

Another benefit of Open Standards is that you can learn about
them by downloading the spec from nice people like Seagate.


Reading exercise
----------------

1) Take 2 minutes to skim the pages from the SCSI manual so you
know what information I have given you.

2) Then read the following questions.

3) Then go in an find the answers to the questions.


Reading questions for the SCSI manual.  Some of these questions
depend on background information you may not have yet, so if
you can't answer some (or any) of them, that's ok (for now)!

1) What's an initiator?




2) What are the four parts of a Command Descriptor Block?




3) If something goes wrong during a request, how does the initiator
   find out about it?




4) What is a logical block address?  How many bits does it contain?
   



5) What does it mean if the transfer length in a request is zero?




6) How does the drive know whether or not a command is linked?




7) In a read command, how many bits is the Logical Block Address?
   How many different blocks can be addressed?




8) How does the initiator know the range of valid block addresses?





9) What happens if a request references an illegal block address?





10) Taking advantage of what we know about disk drive mechanics,
    describe the sequence of steps that occurs between the initiation
    of a read command and the beginning of data transfer.




11) What is prefetching?




12) What effect does the RCD bit have during a write command?