Hard Disk Drives (HDDS)

HDD (software) interface

Modern conventional hard drives export a linear address space. In other words, they present the OS with the illusion of an array of n logical block addresses (LBAs), numbered from 0 to n-1, that the OS can access by block number. The OS must read/write at the granularity of blocks.

Most modern hard drives are “advanced format” HDDs, which means that their sector size exceeds the older standard of 512 bytes. A common physical block size is 4 KiB (4096 bytes). DRAM pages are commonly 4096 bytes as well; this is not a coincidence! There are clear benefits to aligning the caching granularity with device block size.

HDD Guarantees

• Sector writes are atomic: either the entire sector is written, or the previous contents remain unchanged.
  • When a multisector write is only partially completed, it is called a torn write, which is something that software at higher levels must be careful to handle.
• HDD “Unwritten Contract”: accessing two blocks with LBAs that are near to each other is faster than accessing two blocks with LBAs that are far from each other (i.e., sequential reads/writes are faster than random reads/writes)

Disk Parts and Geometry

Disks are essentially a collection of spinning metal surfaces, and as they spin, an arm moves from the center to the outer edge to read and write data.

1. Draw and label a 2D top-down view of a disk. Draw your disk with 1 disk arm, 3 tracks, and 8 sectors per track. (Place LBA 0 on the outermost track, and let the disk rotate counter-clockwise)

Looking at your picture above, does each sector occupy the same area on the platter? Since the platter spins as a unit, how fast does a point located on the outside of the platter move relative to a point located on the inside of the platter?

Breaking down an I/O

To read or write data from the platter, the disk arm must pass over the target sector. There are two parts to any I/O: * Setup * Transfer

We can think of the setup as all the work necessary to locate the disk arm at the appropriate physical location, and the transfer as all of the work that takes place starting from the moment that the reading/writing of the first byte begins.

1. The two components of the setup cost are the seek and the rotational delay. What physical actions correspond to each disk component in your diagram above?
2. On average, how many rotations does a given I/O’s setup phase require?
3. Suppose that the disk spins at 7,200 RPM. How much time is a single rotation?
4. Give ballpark (order of magnitude) costs (latency, in ms) for a commodity HDD to complete

• one disk seek:
• one disk rotation:

Other Physical HDD Considerations

Geometry

Disk geometry affects the costs of I/Os in interesting ways.

1. What is meant by track skew, and what problem is it designed to solve?
2. Why are many conventional hard disks divided into zones (a zone is a contiguous collection of tracks; any two tracks inside a single zone have the same number of sectors per track, but two tracks from different zones have different numbers of sectors per track)?

Caching

HDDs have their own internal memory buffers (disk buffers are O(10s of MiBs))

1. How might the disk’s internal buffer be used to speed up certain reads (hint: locality)?
2. How might the disk’s internal buffer be used to speed up writes?
3. Are either of the above two optimizations (buffering reads, buffering writes) dangerous?
4. Explain the difference between write-back and write-through caching, and which caching strategy makes the most sense for the HDD’s internal buffer.

Scheduling

If the OS and disk were to attempt to immediately satisfy every I/O request, and to strictly complete requests in the order that they were received (i.e. FIFO request scheduling), performance would likely be poor. Why? Multiple applications may simultaneously request access to the disk, and they must all share a single disk head when reading/writing data. Further, no single application’s requests will exhibit perfect locality (metadata is often separated from the data it describes, so even sequential data requests may require seeking: the file system must first read the metadata required in order to find the location of the target data). Given these challenges, we need more sophisticated strategies to order requests; we want to avoid as many unnecessary disk seeks as possible. A good scheduling algorithm should strive to:

• Maximize overall system throughput (the system as a whole should make the most possible progress in a given time window), and
• Maximize fairness (no one application should suffer from starvation while other applications make progress)

The textbook describes the principle of shortest job first (SJF). As the scheduler receives a series of requests, a queue may build. The queue can be (re)ordered in a variety of ways.

1. What is the relationship between shortest-seek-time-first (SSTF) and nearest-block-first (NBF)?
2. How does the “elevator” (or SCAN or F-SCAN or C-SCAN or …) algorithm relate to both of the goals above (e.g., maximize system throughput and maximize system fairness)?

The anticipatory scheduler is alluded to in the text: an anticipatory scheduler will pause for a short time before dispatching an I/O from the queue. If a new I/O arrives, the scheduler has more information to help it optimize its I/O ordering.

3. Deliberately slowing down seems like a counter-intuitive strategy, but it works well for disks. Why might the anticipatory scheduler be a better strategy for a HDD than for a faster (lower latency / higher throughput) device?

More complex scheduling algorithms may prioritize certain applications over others.

4. What entity should be responsible for deciding an application’s priority? What privileges should be necessary for an application to declare that its task is more important than others?