Lecture 23 -- I/O, busses/switchfabrics, interrupts/DMA, etc.
==========

Supporting materials
--------------------
- Modern PC Architecture: Figure 8.11 from pg. 585, P&H 3rd ed
- Figure 8.4, pg. 647 from book (structure of a disk), P&H 2nd ed

Administrative
--------------
- hand out final project

What is I/O?
------------
- from/to other devices
- Input: keyboards, mice, etc.
- Output: display
- Input/Output: disk!, network controller!

Key characteristics
-------------------
- communication of data
- flexiblity in configuring computer (by system builder or end user)
- desire for multiple chips to control costs
    (die area and pin count issues)

Why does it matter?
-------------------
- when was the last time you felt your computer was "slow"?
- performance often limited by I/O, not processor:
  . start-up of computer or applications -- mostly disk I/O
  . compiling software -- often limited by disk I/O, esp for short
    source files and low optimization levels
  . web surfing -- mostly network I/O
  . 'server' applications -- credit-card transactions, airline systems, ...
- device hardware and operating system have biggest impact
  . e.g. scheduling disk accesses
  . e.g. zero-copy TCP/IP stacks
- but I/O architecture of system matters too
  . especially important for 3D graphics
  . also matters for disks and network

What is the I/O architecture in a modern PC?
--------------------------------------------
* Modern PC Architecture: Figure 8.11 from pg. 585
   . North bridge -- high throughput
   . South bridge -- pin expander
* Contrast Intel with AMD architectures
   . Tension between cost, flexibility, performance, ...
* Busses in the machine
  - "front-side bus" (intel only)
  - memory bus
  - USB
  - PCI/AGP, PCI express
  - ATA, Serial ATA
* Built-in capabilities:
  - Wireless network (often on chipset)
  - Sound
  - Old-style serial, parallel

Key points
----------
- I/O benchmarks -- TPC-C for example, WinBench???
- Polling, interrupts, DMA
- Throughput vs. latency...
  Note that taking advantage of high throughput requires lots of transactions
  Also: Thoughput can be measured as bytes/second *or* transactions/second!!
- Interconnect fabric: Bus vs. Network

Disks
-----
- Figure 8.4 from page 657 of book (2nd ed)
- Read/write head...
- Seek time of ~ 8.3 msec
- How many seeks/second?

Networking
----------
- 1Gbit/sec = 100 MBytes/sec --> 30 processor clocks/byte @ 3GHz processor clk

Local interconnection networks
------------------------------
- Busses vs. Switch fabrics
- PCI vs. PCI Express

Busses
------
* Why do you want them?
    . multiple chips
    . flexibility by system manufacterer
    . flexibility by end user
- figure 8.9c, pg. 659
- Synchronous vs. asychronous -- is there a clock signal?
  USB is asychronous
- Asynchronous -- figure 8.10
- Bus protocols:
  . only one 'talker' at once:  need to 'acquire' the bus
  . how to 'arbitrate' on multiple near-simultaneous acquisition attempts?
  . control signals vs. address signals vs. data signals
    separate wires or shared wires?
- Physical/electrical limitations: maximum bus length!

Switch fabrics
--------------
- how to route?
- blocking or non-blocking?
- packet or circuit switched?
- if you have one hub, these decisions can be simpler

Tradeoffs in interconnect design
--------------------------------
* distance vs. speed
* bandwidth vs. cost
* speed vs. robustness (e.g. synchronous vs. asynchronous)

Processor interface
-------------------
- memory mapped I/O vs. separate I/O address space?
- polling... sucks
- interrups -- better, but w/high data xfer rates still slow
- DMA -- I/O controller is effectively a separate processor...
    cache coherence issues.
    still need IRQ to signal when done.
- close interplay between processor's I/O instructions/architecture;
  overall system I/O architecture;
  and operating system.