This note covers the hardware side of I/O (controllers, buses, DMA), what the OS’s I/O system has to provide, and the characteristics of secondary storage. It is marked needs-review because the storage prices below came from lecture without a source.
I/O hardware
I/O devices are typically block devices, which transfer data in fixed-size blocks, or character devices, which transfer data one character at a time as a stream.
A device controller is the hardware that connects the CPU to a device, and it is a small computer in its own right. It sends commands to the device and moves data in both directions. The CPU talks to controllers through controller registers or memory-mapped I/O, or hands transfers off to direct memory access (DMA).
Old computers hung everything off a single bus (the system bus) connecting the CPU, memory, and I/O devices, a topology like old Ethernet networks with a single broadcast domain. Modern systems use multiple buses. The PCI bus is a high speed backbone, and the other buses (memory, SCSI, USB, and so on) branch off of it.
The I/O system has to cope with a wide variety of devices that differ in transfer rate, data format, and control mechanism.
What the OS I/O system provides
- A uniform interface across many devices, plus device-specific interfaces where necessary.
- Device communication and interaction through device drivers.
- A unique ID per device so applications can refer to it.
Ways to drive a device
- Programmed I/O with polling: the CPU issues an I/O command for the process, and the process busy-waits until the I/O completes. Inefficient, since the CPU is tied up the whole time.
- Interrupt-driven I/O: the CPU issues the command and keeps executing. When the I/O completes, the device interrupts the CPU, which then handles the completion. The process may or may not block while waiting, but the processor stays free.
- Direct Memory Access (DMA): the DMA module moves a block of data between the I/O module and main memory using physical addresses. The processor requests the transfer and gets interrupted when it finishes, so it never has to touch the data in between.
What each step buys
Each option down the list frees more CPU: polling burns it for the whole transfer, interrupts free it between transfers, and DMA frees it during them.
sequenceDiagram participant P as Process participant OS as OS driver participant D as Controller + DMA P->>OS: read request OS->>D: issue command, set up DMA transfer Note over P,OS: process blocks, CPU runs other work D->>D: move block directly into main memory D-->>OS: interrupt on completion OS-->>P: wake process, return data
Secondary storage
Everything outside primary memory (RAM) counts as secondary storage: hard drives, SSDs, and other storage devices. Secondary storage doesn’t allow direct execution of instructions or data access via load/store instructions; access goes through I/O operations.
Secondary storage is non-volatile, so data survives power loss. It is very slow compared to primary storage, it is failure-prone, and it is enormous for the price. Rough 2024 street prices from lecture:
- 2 TB HDD for 0.04/GB
- 30 TB HDD for $700
- 500 GB SSD for 0.10/GB
- 100 TB SSD for $40,000
Memory hierarchy
| Level | Speed | Cost | Size | Volatility |
|---|---|---|---|---|
| Registers | Fastest | Most Expensive | Smallest | Volatile |
| L1 Cache | Fast | Expensive | Small | Volatile |
| L2 Cache | Fast | Less Expensive | Still not a lot | Volatile |
| Main Memory | Slower | Less Expensive | Larger | Volatile |
| Secondary Storage | Slow | Cheap | Largest | Non-Volatile |
| Tertiary Storage | Slowest | Least Expensive | Largest | Non-Volatile |
HDDs and SSDs
HDDs are mechanical devices that store data on spinning disks, with a read/write head that seeks across the disk. They are slow, but capacity is large and cheap. SSDs store data in flash memory. They are much faster than HDDs and cost more per byte. They are also more reliable and draw less power.
Disks and the OS
Disks are messy, slow, error-prone devices, and it’s the OS’s job to make them look clean, fast, and easy to use.
The OS typically exposes disk access at different levels to different clients:
- Physical block access: read and write blocks at physical locations on the disk.
- Logical block access: read and write by disk block number, without knowing the physical location of the block.
- File system: read and write files at a specified offset, block, or byte.
Old disks only offered physical block access. Modern disk controllers map the physical geometry (cylinders, sectors, and so on) to logical block numbers through , so the OS sees one contiguous range of blocks.
Performance
An HDD’s performance is dominated by its mechanically moving parts. Limiting seeks helps, and defragmenting helps, but only to an extent.