Finish Cache CSIT 301 Blum 1 Cache Misses

Finish Cache CSIT 301 (Blum) 1

Cache Misses • When a cache miss occurs, several factors have to be considered. For example, – We want the new memory location written into the cache, but where? – Can we continue attempting other cache interactions or should we wait? – What if the cached data has been modified? – Should we do anything with the data we are taking out of the cache? CSIT 301 (Blum) 2

Replacement Policy • Upon a cache miss, the memory that was not found in cache will be written to cache, but where? – In Direct Mapping, there is no choice it can only be written to the cache address it is mapped to. – In Associative and Set-Associative there is a choice in what to replace. CSIT 301 (Blum) 3

Replacement Policy (Cont. ) • Least Recently Used (LRU) – One approach is to track the use of what is in cache and to replace the line that is being used the least. – This is best in keeping with the locality of reference notion behind cache, but it requires a fair amount of overhead. • Not-Most-Recently Used – Another approach is to choose a line at random except that one protects the line (from the set) that has been used most recently – Less overhead CSIT 301 (Blum) 4

Blocking or Non-Blocking Cache • Replacement requires interacting with a slower type of memory (a lower level of cache or main memory). Do we allow the processor to continue to access cache during this procedure or not? • This is the distinction between blocking and nonblocking cache. – In blocking, all cache transactions must wait until the cache has been updated. – In non-blocking, other cache transactions are possible. CSIT 301 (Blum) 5

Cache Write Policy • The data cache may not only be read but may also be written to. But cache is just standing in as a handy representative for main memory. That’s really where one wants to write. This is relatively slow just as reading from main memory is relatively slow. • Rules about when one does this writing to memory is called one’s write policy. • One reason for separating data cache and instruction cache is that the instruction cache does not require a write policy. – Recall the separation is known as the Harvard cache. CSIT 301 (Blum) 6

Write-Back Cache • Because writing to memory is slow, in Write-Back Cache, a. k. a. "copy back” cache, one waits to until the line of cache is being replaced to write any values back to memory. – Main memory and cache are inconsistent but the cache value will always be used. – In such a case the memory is said to be “stale. ” CSIT 301 (Blum) 7

Dirty Bit • Since writing back to main memory is slow, one only wants to do it if necessary, that is, if some part of line has been updated. • Each line of cache has a “dirty bit” which tells the cache controller whether or not the line has been updated since it was last replaced. – Only if the dirty bit is flipped does one need to write back. CSIT 301 (Blum) 8

Pros and Cons of Write Back • Pro: Write Back takes advantage of the locality of reference concept. If the line of cache is written to, it’s likely to be written to again soon (before it is replaced). • Con: When one writes back to main memory, one must write the entire line. CSIT 301 (Blum) 9

Write-Through Cache • With Write-Through Cache, one writes the value back to memory every time a cache line is updated. – Con: Effectively a write-through cache is being used as a cache (fast stand in for memory) only for purposes of reading and not for writing. – Pro: When one writes, one is only writing a byte instead of a line. That’s not much of an advantage given the efficiency of burst/page reading/writing when the cache interacts with memory. – Pro: Integrity: cache and memory always agree CSIT 301 (Blum) 10

Comparing Policies • Write back is more efficient. • Write through maintains integrity. • Integrity is not so much an issue at the SRAM -DRAM interface in the memory hierarchy since both are volatile. – This issue is more important at the next lower interface main memory/virtual memory as virtual memory is non-volatile. CSIT 301 (Blum) 11

Victim Cache • Other than write modified data back to memory, what do we do with the data that is being replaced? • One answer is nothing. • Another possibility is to store it in a buffer that is faster than the next lower level, effectively introducing another small level of cache. This is as the victim cache or victim buffer. CSIT 301 (Blum) 12

References • Computer Architecture, Nicholas Carter • http: //www. simmtester. com/page/news/sho wpubnews. asp? num=101 • http: //www. pcguide. com/ref/mbsys/cache/ • http: //www. howstuffworks. com/cache. htm/ printable • http: //slcentral. com/articles/00/10/cache/pri nt. php CSIT 301 (Blum) 13

Disk Drives CSIT 301 (Blum) 14

Hard drive • In our original view of a computer as being comprised of ALU, Control, Memory, Input and Output, the hard drive is a device connected either as input or output. • But the hard drive is also sometimes viewed as a logical extension of memory. • The hard drive is the primary storage device. – Compared to RAM, storage is non-volatile – Compared to ROM, storage is more easily written. CSIT 301 (Blum) 15

Speed and Capacity • The two main characteristics of a hard drive are – Its capacity: how much data can it hold – Its speed: how quickly can it be read from or written to • Data intensive applications such as databases, graphics and so on will require pages to swapped in and out of memory. The speed of the hard drive will be an important factor for determining how efficiently such programs run. CSIT 301 (Blum) 16

Analog of Moore • Recall that Moore’s Law concerns the exponential growth of the number of transistors on an integrated chip. • There has been similar exponential growth in the capacity of hard drives. CSIT 301 (Blum) 17

Exponential Improvement in Hard Drive Capacity CSIT 301 (Blum) 18

Paper Tape and Cards • Prior to hard drives, programs and data could be stored on cards or paper tape. • In both cases a hole could correspond to a 1 and the absence of a hole to a 0. CSIT 301 (Blum) 19

Magnetic Tape • Magnetic tape was an improvement upon punch cards and paper tape, both in terms of speed and capacity. • In magnetic tape, a long thin piece of plastic is covered with a ferromagnetic material, such as Ferric oxide (Fe 2 O 3). CSIT 301 (Blum) 20

Magnetic Tapes versus Hard drives • The writing or reading of a individual bit is similar whether we are talking about a hard drive or a magnetic tape. • The difference is one of addressing: – Magnetic tapes have sequential access – Hard drives have random access CSIT 301 (Blum) 21

Magnets • Little magnets align themselves in a particular way when in the presence of a magnetic field – E. g. a compass points North because it is a magnet aligning itself with the Earth’s magnetic field CSIT 301 (Blum) 22

Ferromagnetic • Many atoms are like tiny little magnets, but the little magnets point in random directions and tend to cancel out any magnetic effect on a large scale. • An external magnetic field can make these little magnets line up and produce a large-scale (macroscopic) effect. • If the little magnets remain aligned even when the external magnetic field is removed, then the material is said to be ferromagnetic. • The lining up of the magnets is called magnetization. • A ferromagnetic material “holds” or “remembers” its magnetization state. CSIT 301 (Blum) 23

Writing/Recording • The signal (data changing over time) is fed to a magnet which magnetizes the material on the region of the tape that corresponds to that time. – The tape may record analog or digital information. CSIT 301 (Blum) 24

Reading/Playing • The magnetized region produces a magnetic field of its own. • Any device that can sense this magnetic field can read the information encoded on the tape. • It is important to note that the reading device (head) does not have to be in physical contact with the tape in order to sense the tape region’s magnetic field – just in its vicinity. CSIT 301 (Blum) 25

Floating height/Flying height • As opposed to floppies, VCR and cassette tapes, hard disk heads do not come in contact with the medium they are reading from or writing to. • The distance the head is from the material is one of the important design parameters and is known as the floating height or flying height. CSIT 301 (Blum) 26

Reading, Sensitivity and Density • As the heads are made sensitive to smaller magnetic fields, the region corresponding to a unit of information can be made smaller. • Also as the heads are made to approach the material more closely (where the field is stronger) without touching it, the region corresponding to a unit of information can be made smaller. • If the regions grow smaller, the number of such regions per area (the density) increases. – There are other ways to measure density, so this version is sometimes called the areal density. CSIT 301 (Blum) 27

Exponential Improvement in Hard Drive Density The units for areal density are bits per square inch (BPSI) or in this case MBPSI (mega) CSIT 301 (Blum) 28

Shrinking Bit Size CSIT 301 (Blum) 29

IBM RAMAC • To get a sense of this improvement, consider an early disk drive. • One of the first commercially available hard disks was IBM's RAMAC (Random Access Method of Accounting and Control) introduced in 1956. – – Capacity: about 5 MB Used 50 24" disks!!!! Areal density: 2, 000 bits per square inch Data throughput: 8, 800 bits/s. CSIT 301 (Blum) 30

Form Factor • The areal density improvement has allowed the capacity to increase while the size has decreased. • Drive’s form factors (basically their width and height) have continued to grow smaller and smaller. • The 5. 25 -inch width was a standard. It came in three standard heights – Full-height: 3. 25 inch – Half-height: 1. 625 inch – Third height: 1 inch CSIT 301 (Blum) 31

3. 5 inch Form Factor • But now the 3. 5 -inch width has replaced it as the standard in PCs. • This width comes in two standard heights – Third height: 1 -inch which is standard (slim-line) – Half-height: 1. 625 -inch which is used for higher capacity drives • Besides the overall benefits of miniaturization, smaller widths allow the platters to spin faster and thus help improve the speed. CSIT 301 (Blum) 32

Platters and Spindle Speed • Instead of the long plastic strip of magnetic tape, hard disks have a collection of circular shaped aluminum or glass platters (which serve as the substrate) that are covered with magnetic material (the media layer). • Data is accessed by having the head float over the platter as its spins. – Access speed is thus related to rotational or spindle speed which is measured in RPM (revolutions per minute). – Spindle speeds in the thousands to tens of thousands are typical these days. CSIT 301 (Blum) 33

Hard versus floppy disk • The disks in a hard drive are fairly rigid and hence the name “hard” disk. • The disk in a floppy disk is made of a more flexible material. CSIT 301 (Blum) 34

CSIT 301 (Blum) 35

Anatomy of a Hard Drive • The data is stored on platters: hard, flat circularshaped piece of aluminum coated with magnetic material on one or both sides. • The spindle serves as a rotational axis for the platters. • The rotational motion is driven by the spindle motor. • The information is accessed by a read/write head. – Typically there are two heads per platter, one on the top, one on the bottom. CSIT 301 (Blum) 36

Anatomy of a Hard Drive (Cont. ) • The head is attached to a part called the slider, which is in turn attached to the actuator arm, which is used to position the head in the desired region. • The position of the actuator arm is controlled by the actuator. • There is actually a parallel array of heads which are moved in unison by the actuator. • The actuator in controlled by the logic board which communicates the rest of the PC. CSIT 301 (Blum) 37

CSIT 301 (Blum) 38

Don’t try this at home • Reading requires the head to come very, very close to the platter without touching it. • Hard drives should not be opened because a typical speck of dust is larger than the headto-platter reading distance. • Opening a drive will almost assuredly ruin it. – Head crash: when the head touches the platter CSIT 301 (Blum) 39

Dust particle versus Flying Height CSIT 301 (Blum) 40

Tracks • In both hard and floppy disks, the data is written in concentric circular paths known as tracks • A typical floppy disk has 80 (doubledensity) or 160 (highdensity) tracks. CSIT 301 (Blum) track 41

Tracks (cont. ) • The density of tracks is measured in units of tracks per inch (TPI). • Each track is further divided into sectors. • The location of information is remembered by noting its track and sector numbers. CSIT 301 (Blum) 42

Sectors • Radial lines break the tracks up into sectors, each of which holds 512 bytes of information. Sector Usually 17 sectors per track for a floppy CSIT 301 (Blum) 43

Not all sectors are created equal • A sector on the outer portion of the platter has a greater area than a sector on the inner portion. – More and more of the storage area is wasted as one moves out in the radial direction. – More modern drive technologies divide the outer tracks into more sectors to make use of this storage area. CSIT 301 (Blum) 44

Zoned Bit Recording • Zoned bit recording (ZBR), a. k. a. multiple zone recording or zone recording. • Tracks are broken into groups called zones based on their radial position. Tracks with greater radii are broken into more sectors so that storage area is not wasted. – Requires more sophisticated controller. CSIT 301 (Blum) 45

Larger Radius, Faster Access • In ZBR, the outer tracks have more sectors and thus hold more data, but the data is accessed by the spinning of the disk. So more data goes by per revolution when reading the outer tracks. • The outer tracks tend to be used first. So disk access performance may go down as one starts to use the inner/slower tracks. CSIT 301 (Blum) 46

ZBR CSIT 301 (Blum) 47

Cylinders • Each platter has tracks with the various radii. • All the tracks on different platters but with the same radius make up a cylinder. • For example, if a hard drive has – Four platters – And each platter has 600 tracks • Then – There will be 600 cylinders – Each cylinder will have 8 tracks (assuming that each platter has tracks on both sides). CSIT 301 (Blum) 48

Why cylinders are important • If the data being written does not fit in a single track, it can be spread across the cylinder. • Activating different heads (at the same radius/cylinder) is an electronic process and thus is faster than moving the arm to a new radius, which is a mechanical process and thus slow. CSIT 301 (Blum) 49

Accessing a disk 1. The combination of application, operating system, BIOS and possibly disk interface circuitry determine the location of the data on the hard disk. 2. The location corresponds to a geometric location on the disk. One needs to know A. The cylinder which determines the radius or track. B. The head which determines which platter and which side of the platter. C. The sector which determines where along the track. CSIT 301 (Blum) 50

Accessing a disk (Cont. ) 3. See if the information is cached before starting the slower process of accessing it on the drive. 4. If the drive is not already spinning, it must be “spun up” to its working rotational speed. It might have been “spun down” to save energy. 5. The actuator moves the heads to the appropriate cylinder (track, radial position). 6. The actuator then selects the appropriate head and waits until the selected sector passes by the head. It then reads. CSIT 301 (Blum) 51

Accessing a disk (Cont. ) 7. The information is read and placed temporarily into a buffer. 8. The hard disk interface then sends the data to some other part of the PC, in most cases the memory. CSIT 301 (Blum) 52

Cache on the Hard Disk • The hard disk has a cache/buffer. – After requesting data from the hard drive, one does not sit by idly waiting for the result. One carries on with whatever else possible. The result of the read is placed in a buffer and picked up when the processor/memory is ready to take it. – When reading, one also grabs the data in the neighboring sectors and places it in the buffer (prefetching). This is the standard idea of caching something you expect to need soon based on locality of reference. CSIT 301 (Blum) 53

Size and Distinction • The cache on the hard disk is typically between 512 KB and 2 MB. – Some SCSI drives may have as much as 16 MB. • Be careful not to confuse the cache on the hard drive with “disk cache” which refers to a section of main memory used to hold data recently read from the hard drive. CSIT 301 (Blum) 54

Seek and ye shall find, but be quick about it • Seek time is the time required to position the head to the selected cylinder. – Typical seek times are in milliseconds (ms) – Recall that processor times are in nanoseconds (ns, approx. a million times smaller) and memory times are in microseconds ( s, approx. a thousand times smaller). • Seek time is not access time, but is probably the major part thereof. CSIT 301 (Blum) 55

Versions of Seek Time • Average: from a random track to another random track – This is what is typically reported – 8 – 10 ms • Track-to-track: from one track to the adjacent track • Full stroke: the full range from the innermost to outermost track CSIT 301 (Blum) 56

Settle Time • After the head has reached the appropriate cylinder (track, radius), it must take a short amount of time to stabilize before reading can occur. • This time is called the settle time or settling time. • It is short compared to seek time and does not vary much from manufacturer to manufacturer. CSIT 301 (Blum) 57

Command overhead time • Command overhead time is the time it takes from when the hard drive is given the read instruction to when the actuator starts positioning the head – Since this is an electronic time as opposed to the mechanical seek time, it is much smaller and does not contribute much to the access time. CSIT 301 (Blum) 58

Latency • After the instructions have been interpreted and the actuator begins to move (command overhead) and the head has reached the selected cylinder (seek) and has stabilized (settle), it is still not ready to read. • It must wait until the selected sector rotates past the head. This time is called the latency. CSIT 301 (Blum) 59

Latency and Spindle Speed • The time it must wait for the correct sector to swing by clearly depends on how fast the disks are rotating – the spindle speed. – If the spindles rotates at 10, 000 RPM (revolutions per minute), then it rotates at speed of 10, 000/60 = 166. 7 revolutions per second. – If there are 166. 7 revolutions per second, then a revolution takes 1/166. 7 seconds = 0. 006 s or 6 ms. – The average latency is half of the rotation time or in this case 3 ms. CSIT 301 (Blum) 60

PCGuide table CSIT 301 (Blum) 61

Ordering the to-do list • Because the hard drive is slower than the processor and memory, there may be a back up of tasks for it to perform. The order in which it performs these tasks can greatly affect its efficiency. • One ordering is a simple FIFO (first-in, first-out) ordering. The tasks (reads and writes) are queued up and the first task requested is the first task performed. CSIT 301 (Blum) 62

More Sophisticated Orderings • Seek-Time Optimization (a. k. a. Elevator seeking): – Seek time involves the radial positioning of the head. The tasks are ordered based on their radial positioning to minimize seek time. • Access-Time Optimization (a. k. a. multiple command reordering): – Takes into account both radial and angular positioning to minimize access time which includes seek time and latency. CSIT 301 (Blum) 63

Ordering Comparison CSIT 301 (Blum) 64

Locating data • The platters are dense with data, if the head is even the slightest bit off target, one will be reading the wrong data. • The actuator uses a voice coil. The voice coil uses a feedback mechanism to locate the data. – Feedback means taking some output and putting it back in as input. In this case the head is reading information from the platter. Some of that information is telling the head where it is on the platter. CSIT 301 (Blum) 65

Servo Information • The information on the disk about location on the disk is called servo information. – It is not user’s data but data about the disk’s operation. • The servo information can be placed – In dedicated sectors (wedge servo) – On dedicated platter side (dedicated servo) – Spread throughout (embedded servo) CSIT 301 (Blum) 66

CSIT 301 (Blum) 67

Low-level format • Low-level formatting (LLF), a. k. a. physical formatting, establishes the tracks, sectors, etc. and writes the servo information. • For a hard drive, this is done once by the manufacturer. • For a floppy disk, this can be done by the user. • Any previous data is lost when a low-level format is performed on a disk. CSIT 301 (Blum) 68

Our old enemy heat • When objects heat up, they tend to expand (thermal expansion). • If the heat in a hard drive is not distributed uniformly, then different platters are at different temperatures and thus have expanded different amounts. • This is why dedicated servo is not as good as embedded servo. CSIT 301 (Blum) 69

Thermal recalibration • Every once in awhile the hard disk reads itself just to check on track positioning and so forth. – You may hear the hard drive spinning even though the PC is neither reading it nor writing to it. – This is called thermal recalibration. – The amount of thermal calibration necessary was reduced by manufacturers moving to embedded servo techniques. CSIT 301 (Blum) 70

A gentle breeze • While reading or writing the platters spin, which causes a breeze. The actuator arm is designed like a wing that floats in this breeze. • When the drive spins down (stops spinning), the breeze stops blowing and the actuator arm must come in for a landing. It comes in contact with the platter. • We do not want there to be any data written in the region where the arm lands. CSIT 301 (Blum) 71

The landing zone is for … • There are special regions on the disk called landing zones where no data is written and where the head comes to rest when the drive spins down. • Positioning the head into the landing zone is called head parking. • There is a BIOS parameter that informs the system where the landing zones are. But in modern drives the drive is designed to return automatically to the landing zone (even if the power is lost). • Some drives are designed so that the head never lands. CSIT 301 (Blum) 72

Landing Zone Setting in BIOS CSIT 301 (Blum) 73

Spindle Speed Revisited • In addition to accurately positioning the head, the platter must spin at a very accurate speed. • The standard spindle speed for a long time was 3600 RPM (revolutions per minutes), which corresponds to 60 cycles per second or 60 Hz, which is same frequency as the AC power that comes out of a standard US wall socket. • But since latency is connected to spindle speed, the spindle speeds have been raised. • The higher speeds tend to debut with SCSI drives and then filter down to ATA/IDE drives. CSIT 301 (Blum) 74

Spindle speed and drive type Source: PCGuide. com CSIT 301 (Blum) 75

Spindle Speed Issues • Smaller platters are easier to spin faster – Think of an ice skater spinning • Platters must spin together, so it is easier to spin fewer platters. • The platter must spin at a steady speed and not vibrate (we don’t want a head crash). • Heat and power: spinning faster requires more power and generates more heat. CSIT 301 (Blum) 76

Power Issue • Because of momentum (objects in motion tend to remain in motion), it takes more power to start the drive spinning than to keep it spinning. – This is one place where the 12 V connection is used – “Power management” reduces the power when a device is not being used. But spinning up a hard drive requires more power than maintaining the rotation, so continually starting and stopping could cost power. • Also the starting and stopping could be detrimental to the drive. CSIT 301 (Blum) 77

Power Issue (Cont. ) • When a computer has multiple drives, e. g. a master/slave set up, the slave may delay its spin up until the master has completed its spin up so they are not pulling on the power simultaneously. – One may have to use a particular jumper setting to ensure this delay. CSIT 301 (Blum) 78

Cover and Base Casting Don’t open the case (unless your room is clean). The screws in the case are unusual to prevent you from opening it. CSIT 301 (Blum) 79

Yes to Air / No to Dust • The hard drive cannot be completely sealed off, it needs air to keep it from overheating. • On the other hand, the slightest spec of dust can cause a head crash. • A hard drive needs an air filtration and circulation system. – The circulation occurs naturally in that the spinning disks cause the air to flow. – There is a filter to keep dirt out. It does not need changing. CSIT 301 (Blum) 80

Connecting a Hard Drive • There are typically three sets of pins and/or connectors associated with a hard drive – Power (Molex): the juice that allows it to operate – Data Interface: where data, addresses, instructions enter and leave – Jumper: switches that specify in what mode the drive will operate CSIT 301 (Blum) 81

Power, Jumper and Data Interface CSIT 301 (Blum) 82

References • PC Hardware in a Nutshell, Thompson and Thompson • http: //www. pcguide. com • All-in-One A+ Certification, Meyers and Jernigan CSIT 301 (Blum) 83