Anatomy of storage performance

[DougLite]

U N D E R - C O N S T R U C T I O N

Designing a [H]ard|Drive is a difficult balancing act. The major storage makers must all balance cost, reliability, environmental factors, and performance. One bad design can spell doom for a storage maker, witness IBM's fire sale of their HDD division to Hitachi after the DeskStar 60GXP debacle.

There are three main measures of storage performance: seek performance (how quickly the drive can find a random request), capacity (how much the drive can hold) and transfer performance (how quickly the drive can read or write in a straight transfer with no seeks). Additionally, makers must keep noise and power consumption/heat dissipation in check. In this thread, we will explore considerations and limitations on hard drive design, and how the various considerations often compete with each other. The makers of HDDs must balance all considerations, leading to some compromises that may seem bizarre or poorly thought out at first. However, understanding a hard drive as a complete system with several moving parts that interact with each other gives much insight into the considerations of hard drive design. Also, ATA drive makers use buffer optimizations to mask the mechanical inferiority of 7200RPM drives to 10K and 15K RPM SCSI Drives, that will be discussed later.

[...]

Platters:
What it affects: Capacity, transfer performance
A HDD has at least one rotating disk with data stored on it by manipulating magnetic fields on the surface of that disk. The physical size/surface area of a drive's platter is pretty much fixed, as all desktop drives in this day in age ship in a 3.5" wide x 1" high form factor. However, there are quite a few things a manufacturer can do with platters nonetheless.
Increase the number of platters Seems obvious - the more platters a drive has, the more data it holds, increasing capacity. However, increasing the number of platters requires a more powerful (and power hungry) spindle motor, and increases the power and precision requirements of the actuator that performs seeks. Positioning an actuator on tracks on a single platter is difficult enough, doing so on five (flagship Hitachi drives) or even the more conservative three or four used by other makers, is even more difficult, as the read/write heads must be accurately positioned on all of the platter(s) before a transfer can begin. Since no two platters are exactly alike, drives with higher platter counts often require more time to accurately position the read/write heads decreasing seek performance
Increase the amount of data on each platter This (referred to as increasing density) is generally accepted as the best way to increase capacity, as it does not involve compromising on environmental factors, and is also typically the cheapest way to increase capacity. Makers can increase the number of tracks on a platter, and they can also increase the amount of data in a given track. Both increase capacity, but the latter also increases transfer performance, as the read/write heads pass over more data during a rotation before having to move to a different track. However, increasing density does have drawbacks. Since the data is packed closer together, the magnetic field that represents the data cannot be as strong, or it will affect nearby data. This means that the read/write heads must be more sensitive and precise, and this can lead to increased cost and/or decreased seek performance.

Spindle Motor:
What it affects: Seek performance, transfer performance
This is fairly straight forward - making the motor spin faster increases seek performance by reducing rotational latency, the period of time between when the read/write heads are positioned over the requested track and when the requested sector passes underneath the read/write heads for retrieval. It also increases transfer performance as the faster rotation allows more sectors to pass underneath the heads during a given time frame. However, increasing spindle speed means more power, heat and noise. Also, higher spindle speeds require either raising cost or compromising on reliability, neither of which are good ideas.

Actuator:
What it affects: Seek performance
The actuator is repsonsible for positioning the heads over the appropriate track on the disk to perform read and write operations. There are two goals in its design, power and precision.
Power: The more powerful the actuator, the faster it will move large distances across the platters. Increasing actuator power increases seek performance but requires a higher current draw to move it more forcefully. Also, as the actuator moves faster, its noise profile increases noticably as well.
Precision: The more precise the actuator, the quicker it can read the servo data on the platters and minutely adjust its position to read and write on individual tracks, as the actuator must be accurately positioned over the appropriate track to prevent nearby tracks from affecting read/write operations. As density continues to increase, actuator precision must be increased with it, or the drive's servo system will require multiple cycles to get the heads positioned, leading to decreased seek performance

 

[DougLite]

Buffer Performance:

It is no secret that ATA hard drives are mechanically inferior to SCSI units. SCSI units have as much as twice the spindle speed and half the seek time. ATA drive makers are tightly constrained on cost and environmental factors, making it impossible to deliver quick (around 5ms or below) seeks and high spindle speeds (10,000 or 15,000RPM). They are stuck with 7200RPM spindle speeds and seeks of around 8-10ms, depending on the make and model. The WD Raptor family is an exception to these rules, but noone has followed WD's path with SCSI like ATA drives.

However, the ATA drive makers do have two things they can use to much more effect than a SCSI drive: firmware and buffer. Together, the drive's firmware and buffer work to mask the mechanical limitations of hard drives.

Buffer
Bigger is always better here. The bigger a drive's buffer, the more aggressively it can read ahead, the more writes it can cache, and in general the more it can do before the user or system actually requests it. Western Digital started the buffer race back in 2001 when they upgunned their already performance leading WD1000BB with an 8MB buffer. Prior to the release of the WD1000JB, 2MB buffers were the norm. At present, many makers are shipping flagship drives with 16MB buffers. Generally, both of these moves have boosted desktop storage performance about 20%. There are two reasons why buffers work so well in application, even though their impact does not show up in synthetic measurements of a given drive's mechanical performance: write caching and read ahead algortihms. The principle behind write caching is pretty simple: Rather than taking the time to actually make a time consuming seek and write operation, the drive simply stores the written data in the buffer, then waits for activity to die down and writes the data when it's convenient for hte drive and it is not busy fulfilling other requests. Hence, you may hear or read of write caching being called "lazy write." Read ahead however is more of art form than a science, and that leads us into the discussion on firmware design. The overall goal of firmware design is to read data from the disks as the drive gets to it, rather than when it is actually requested by the user or system.

(To be continued)

 

[unhappy_mage]

Positioning an actuator on tracks on a single platter is difficult enough, doing so on five (flagship Hitachi drives) or even the more conservative three or four used by other makers, is even more difficult, as the read/write heads must be accurately positioned on all of the platter(s) before a transfer can begin. Since no two platters are exactly alike, drives with higher platter counts often require more time to accurately position the read/write heads decreasing seek performance

Wouldn't this be more a product of a higher rotational momentum of a heavier set of heads rather than having to position multiple heads? The heads are on a single pivot - they all point to the same position on their respective platters - so it's not like they're all aligning to different places.

Also, I find the dark-green on dark-grey hard to read; maybe green or even bright green.

 

[DougLite]

In theory yes, they all point to the same point. However, tolerances are REALLY tight, and even the slightest variances can cause problems. Yes higher actuator mass also requires more power to hold the line on seek performance, but precision is also a factor in the maker's decisions on how many platters they will use. There is a reason that Hitachi remains the only maker with five platter drives, and that SCSI flagships stick to four.

Also note that Maxtor specifies variable seek time claims on their flagship Atlas 15K II, depending on the number of platters... 3.0ms for 1 platter units, 3.1 for two, and 3.3 for four.