A System Perspective: Processors, Memory and InputOutput

1
A System Perspective Processors, Memory and
Input/Output

• Last Time
• Interrupt-driven Input/Output
• Direct Memory Access, slave processors, block
  Input/Output
• This Time
• A perspective on the entire computer
• Relative speeds, rates, and relationships
• Reminders/Announcements
• HW 4 due Friday 6/4 (in class)
• Course review on Friday
• Final Exam, 6/7/99, 8-11am
• A - O Center 113
• P - S Center 217A
• T - Z Center 217B

2
Input/Output Perspective

• How does all of this really fit together?
• Bus arbitration, interrupts, and DMA
• Typically, Memory and I/O busses are multi-master
• Arbitration precedes all of the major bus
  activity
• .... lifetime of a computation, or an I/O
  operation....

3
Basic Computation

• Processor computes, accessing its cache, many
  hits and local computation (hopefully)
• Misses in cache, must go out over the memory
  bus.
• 1st arbitrates for the bus
• 2nd accesses the memory
• 3rd memory arbitrates for bus (if split
  transaction, otherwise bus held during memory
  latency)
• 4th data is returned
• Bus arbitration increases Cache miss penalty

4
Basic Input/Output

• Polled I/O -- as before, processor arbitrates for
  bus and mediates everything
• Interrupt-driven I/O -- basically as before
• DMA based I/O -- the Direct memory access
  controller must be able to master the bus, to
  transfer the data to memory
• Both processor and DMA controller must arbitrate
  for the bus before using it.
• Both must release the bus in a reasonable amount
  of time to avoid long waits.

5
Entire DMA based I/O Operation

• Processor arbitrates, writes I/O Control regs,
  I/O begins
• I/O interrupts processor, Processor arbitrates,
  writes DMA control regs
• DMA arbitrates, moves data into memory
• DMA interrupts processor, processor arbitrates,
  updates memory structures
• Arbitration buried at many steps -- must be done
  fast.

6
Input/Output Summary

• Processor-mediated Input/Output
• Interrupt-driven Input/Output
• DMA based Input/Output
• Most of this is done in the operating systems
  (kernel and device drivers)
• In a few low-level systems (Macs, PCs), some of
  this can be done by the user (if desired).
• gt Major reason for PC hardware compatibility
• Early MSDOS I/O was implemented poorly (slow)
• Applications bypassed the OS, to get better
  performance
• Now, they depend critically on the underlying
  hardware

7
System Perspective

• Computer elements
• Relative Speeds and Rates
• Where are the bottlenecks?

8
Processing Speeds

• Pipelined Instruction processor
• Built from extremely fast (and small) Silicon
  transistors
• Execute at 500Mhz , multiple issue
• 2x109 instructions per second (billions)
• ... except, sometimes it has to access the memory
  ...

9
Memory Speeds

• Memory speeds varies with size and cost.
• Fastest is equal to the processor speeds.
• Bulk is much slower 100s of megabytes of
  storage
• 50ns typical access time, 20 Millions accesses
  per second
• There must be 100s to 1000s of instructions per
  slow memory access to achieve good performance
• lots of cache hits, system BW of 1GB/s

10
Input/Output Speeds

• I/O devices are generally larger.
• Wide variance in properties, but slower.
• Disks -gt 10ms seeks, 100Hz, we have at least ten
  millions of instruction times for a disk
  input/output operation
• Monitor (frame buffer) -gt 60 Hz, we have 100s of
  millions of instructions between screen refreshes
• Keyboard -gt 10Hz, we have billions of
  instructions per keyboard stroke

11
Input/Output Perspective

• Input/Output Events are usually extremely
  slow/rare. Not glamorous, poor stepchild.
• Handling them efficiently is important to free
  the processor to do other work.
• Problem processors and memories are scaling up
  in performance much faster than input/output
  devices (disks, in particular).
• Many systems / computations are becoming
  input/output bound. Further, many new
  applications are I/O intensive.
• High performance I/O is a hot topic of research!
• parallelism, prefetching, caching, storage
  capacity

12
Relative Time (area)

• Instruction
• Memory Access
• Average Disk Seek (access)

13
Parallel Input/Output
Parallel Disks
Parallel Tapes

• Parallelism for higher bandwidth and lower
  transfer latency
• How to organize data across the parallel units?
• Prefetching (anticipating to reduce latency)
• Caching (same idea as before), but petabytes
  (1015 bytes) of storage
• a terabyte of disk only cost 25K, aggregate
  transfer BW 1GB/s

14
Multi-media Servers

• Audio servers
• Answering machines, Voice manipulation
• Video servers
• Picture phones (answering machines)
• Video editing and general manipulation
• How much data in a single NTSC signal?
• 100Mbits/sec, HDTV? 2-4x
• a 2-hour movie on NTSC?
• 100GB, 2GB compressed
• HDTV?
• 400GB, 8GB compressed
• Data volumes are HUGE, these will be I/O
  intensive systems.

15
Video on Demand (Blockbuster in your dorm room)
Video or Multimedia Server
High Speed Network
Parallel Tapes
Parallel Disks

• How do we build a system to support 100
  simultaneous video streams? 1000?
• What about when every wants to watch at the same
  time?
• What if everyone wants to watch the same movie?

16
Large Scale Web Servers
WWW Server
High Speed Network
Parallel Tapes
Parallel Disks

• Serves multimedia, but perhaps less predictably
  and in smaller chunks
• What do these workloads look like? Hot page,
  fresh news? Free software?
• The next generation of machine design problems
  appear to be I/O intensive and information
  intensive NOT compute intensive...

17
Next time....

• Whirlwind review of the course