Other Storage Devices

Other Storage Devices

A challenger to magnetic disks.
Optical disks (CDs)

Optical disks contain data stored optically rather than magnetically.

Optical storage is removable and inexpensive to manufacture.

For example, it costs $1 to $2 per 500MB CD-ROM.

CDs are a popular medium for the distribution of software.

Optical disks are usually write-once.

They are also called WORM (Write Once, Read Many) storage.

Agreement on standards for CDs have slowed their time-to-market, allowing magnetic disks to stay ahead.

However, writable optical disks may have the potential to compete with new tape technologies for archival storage.

Other Storage Devices

Magnetic tapes

Magnetic tapes use the same basic kind of recording technology as magnetic disks.

The major difference:

For tapes, the read/write heads actually touch magnetic tapes.
For disks, they float above the disk surfaces (the Bernoulli effect).

Tapes are cheaper than disks for two reasons:

They pack more medium into a fixed size by recording on long strips.
They are removable (allowing one reader to access multiple tapes).

This means that a 9GB cartridge can cost less than $20.

$2/GB is considerably less than the $30/GB for magnetic disk.

Other Storage Devices

Magnetic tapes
Longitudinal scan (linear scan)

This is similar to audio cassette tapes.

Data is written in a track that runs the length of the tape.

As with audio tape, a single tape may have more than one track, and the tracks may run in opposite directions.

Helical scan

These tapes put their heads on a rapidly spinning cylinder, which spin at on a diagonal to the tape.

By increasing the speed of the heads relative to the tape, they allow much denser storage (20 to 50 times).

The problem is that they cannot search as quickly, and they wear out more quickly (100's of passes versus >1000's of passes for longitudinal.)

Other Storage Devices

Robots (tape or optical disk)

Robots can be used automatically move media (tapes or optical disks) between drives and shelves.

This allows terabytes of data to be accessible within ten's of seconds.

Since reader cost is fixed, tape robots cost less per unit of storage as more media share the same number of readers

The robot cost is also amortized.

Of course, more tapes per drive means more contention for drives and lower bandwidth per tape.

A large tape robot can cost as little as $40 per GB for a full system.

Other Storage Devices

Holographic storage (future?)

There is lots of research on using cubes of light-sensitive material to store information holographically.

A device would read or write an entire "slice" at once using lasers.

This would provide very high bandwidth and high storage density.

There are currently several problems researchers are working to solve.

In particular, making this medium read/write is difficult.

Buses

The bus serves as a shared communication link between the subsystems of a computer, e.g. the CPU, memory and the I/O devices.

Buses

Bus advantages:
They are low cost.

A single set of wires is used to connect multiple components.

They are versatile.

New devices can be easily added or removed.

Disadvantage:
They create a communication bottleneck , possibly limiting I/O throughput.

Two types:
CPU-memory buses

Short, generally high speed and matched to maximize memory-CPU bandwidth.

I/O buses

Lengthy with many types of devices connected to them, each of which may have different data bandwidths.
They usually follow a standard.

Buses

Basic bus transactions:
If the transaction is a write (to memory), the data may be sent immediately after (or perhaps in parallel with) the address.
For reads (from memory), there is usually a delay between the time the address is sent and the time the data is ready.

This delay occurs because of the memory latency.

Typical bus read transaction (on a synchronous bus):

Bus Design Decisions

Cost versus performance trade-offs:

Option	Better performance	Lower cost
Bus width	Separate address & data lines (adv for writes)	Multiplex address & data lines
Number of bus lines	Wider is faster	Narrower is cheaper
Transfer size	Multiple words have less bus overhead	Single word transfer has simpler logic
Bus masters	Multiple, arbitrated	Single, no arbitration
Split transaction	Yes. Separate request and reply packets get higher bandwidth	No. Continuous connection is cheaper and has lower latency
Clocking	Synchronous	Asynchronous

The choice of using any of the options listed in the first three rows is straightforward.

All give higher performance at more cost.

Bus Options

Bus masters

A bus master is a device on the bus that has the power to initiate a transaction.

The CPU is always a bus master.
I/O devices can also be bus masters.

With multiple masters, an arbitration scheme is required to determine who gets the bus next.

Split transactions

For many buses, a transaction must be performed atomically.

This means the bus is busy from the time the data is request until the time the request is complete.

This may work well for memory, but it can be intolerable for I/O where data may not be ready for 10ms after the request.

Bus Options

Split transactions

In systems that have multiple bus masters, split transactions provide an improvement in bandwidth.

In a split transaction, the read request is sent and the bus is released while the data is prepared.

Once memory has retrieved the data, it arbitrates for the bus and `tags' the data reply for the CPU.

This makes the bus available for other bus masters while memory reads the words from the requested address.

Split transactions can be even more beneficial for I/O buses.

For example, a SCSI bus can have several long disk transactions in progress at the same time.

However, a split transaction may have higher latency than a bus that is held during the complete transaction (due to arbitration.)

Bus Options

Synchronous vs. asynchronous

Synchronous buses require a fixed protocol for addresses and data relative to a global clock that all devices agree upon.

Signals are only considered valid at certain points on the clock waveform (usually the rising or falling edge).

If all devices are equally "fast," synchronous works well and it is inexpensive and fast.

Two disadvantages:

Cheap devices, such as keyboards, must have logic that is able to run under the same fast clock rate used for disks.
Clock-skew problems limit the length of synchronous buses.

CPU-memory buses are typically synchronous.

Bus Options

Synchronous vs. asynchronous

The alternative is asynchronous logic , which uses hardware handshaking to exchange information on the bus (usually an I/O bus).

Typical write transaction on a asynchronous bus:

Asynchronous buses address previous disadvantages are usually slower because of the overhead associated with the handshaking protocol.

Performance Metrics

Throughput

For I/O devices, throughput is usually called I/O bandwidth .

This measures the amount of data a particular I/O device can transfer in a given period of time.

Response time

This is called I/O latency .

This measures the time taken from when a request is placed into a device queue until the time that service is completed.

Note that this may include time spent waiting in a queue as well as the usual service time.

Performance Metrics

Trading off throughput & response time.

For minimum response time , queues should be kept empty so an incoming request can be serviced as soon as possible.

However, maximum bandwidth makes the opposite true.

Try and keep the queue full so the device always has something to do.

Throughput can be improved in many ways.

For example, a designer could add additional disks to a system allowing more requests to be serviced at a given time.
Requests would likely spend less time waiting in queues, which may improve response time as well.

Human-computer interaction

In general, improving performance does not always improve both response time and throughput.

Given that this is true, which is more important ?

In a multiprogramming environment, one might say throughput .
However, user productivity is non-linearly related to response time as shown below.

Interactions between humans and computers are divided into three "types" of time:
Entry time

This is the time to enter a command into the computer.

System response time

The delay between when the command is completed and the full answer is returned.

Think time

The time from the reception of the response until the user begins to enter the next command.

Human-computer interaction

The sum of these times is the transaction time .

One study shows that user productivity is inversely proportional to transaction time.

In particular, people get more productive as the computer reacts more quickly.

In other words, reducing response time actually decreases transaction time by more than just the reduction in response time.

In general, people need less time to think when given a faster response time.