SQL: Queries, Programming, Triggers - PowerPoint PPT Presentation

About This Presentation
Title:

SQL: Queries, Programming, Triggers

Description:

See the notes for information on how the s are organized. – PowerPoint PPT presentation

Number of Views:21
Avg rating:3.0/5.0
Slides: 47
Provided by: RaghuRa6
Category:

less

Transcript and Presenter's Notes

Title: SQL: Queries, Programming, Triggers


1
Database Systems II Secondary Storage
2
The Memory Hierarchy
Tertiary Storage Tape, Network Backup
Swapping, Main-memory DBMSs
Disk
File System
Virtual Memory
3,200 MB/s (DDR-SDRAM _at_200MHz)
300 MB/s (SATA-300)
Disk-Cache (216MB)
Main Memory
6,400 MB/s 12,800 MB/s (DDR2, dual channel,
800MHz)
16 GB/s (64bit_at_2GHz)
CPU
L1/L2-Cache (256KB4MB)
CPU-to-L1-Cache 5 cycles initial latency,
then burst mode
CPU-to-Main-Memory 200 cycles latency
3
The Memory Hierarchy
  • CacheData and instructions in cache when needed
    by CPU. On-board (L1) cache on same chip as CPU,
    L2 cache on separate chip.Capacity 1MB, access
    time a few nanoseconds.
  • Main memoryAll active programs and data need to
    be in main memory. Capacity 1 GB, access time
    10-100 nanoseconds.

4
The Memory Hierarchy
  • Secondary storageSecondary storage is used for
    permanent storage of large amounts of data,
    typically a magnetic disk.Capacity up to 1 TB,
    access time 10 milliseconds.
  • Tertiary storageTo store data collections that
    do not fit onto secondary storage, e.g. magnetic
    tapes or optical disks.Capacity 1 PB, access
    time seconds / minutes.

5
The Memory Hierarchy
  • Trade-offThe larger the capacity of a storage
    device, the slower the access (and vice versa).
  • A volatile storage device forgets its contents
    when power is switched off, a non-volatile device
    remembers its content.
  • Secondary storage and tertiary storage is
    non-volatile, all others are volatile.
  • DBS needs non-volatile (secondary) storage
    devices to store data permanently.

6
The Memory Hierarchy
  • RAM (main memory) for subset of database used by
    current transactions.
  • Disk to store current version of entire database
    (secondary storage).
  • Tapes for archiving older versions of the
    database (tertiary storage).

7
The Memory Hierarchy
  • Typically programs are executed in virtual memory
    of size equal to the address space of the
    processor.
  • Virtual memory is managed by the operating
    system, which keeps the most relevant part in the
    main memory and the rest on disk.
  • A DBS manages the data itself and does not rely
    on the virtual memory.
  • However, main memory DBS do manage their data
    through virtual memory.

8
Moores Law
  • Gordon Moore in 1965 observed that the density of
    integrated circuits (i.e., number of transistors
    per unit) increased at an exponential rate, thus
    roughly doubles every 18 months.
  • Parameters that follow Moores law
  • - number of instructions per second that can be
    exceuted for unit cost,- number of main memory
    bits that can be bought for unit cost,- number
    of bytes on a disk that can be bought for unit
    cost.

9
Moores Law
Number of transistors on an integrated circuit
10
Moores Law
Disk capacity
11
Moores Law
  • But some other important hardware parameters do
    not follow Moores law and grow much slower.
  • Theses are, in particular, - speed of main
    memory access, and- speed of disk access.
  • For example, disk latencies (seek times) have
    almost stagnated for past 5 years.
  • Thus, moving data from one level of the memory
    hierarchy to the next becomes progressively
    larger.

12
Disks
  • Secondary storage device of choice.
  • Data is stored and retrieved in units called disk
    blocks or pages.
  • Main advantage over tapes random access vs.
    sequential access.
  • Unlike RAM, time to retrieve a disk page varies
    depending upon location on disk.
  • Therefore, relative placement of pages on disk
    has major impact on DBMS performance!

13
Disks
  • Disk consists of two main, moving parts disk
    assembly and head assembly.
  • Disk assembly stores information, head assembly
    reads and writes information.

Spindle
Disk head
Platters
14
Disks
  • The platters rotate around central spindle.
  • Upper and lower platter surfaces are covered with
    magnetic material, which is used to store bits.
  • The arm assembly is moved in or out to position a
    head on a desired track.
  • All tracks under heads at the same time make a
    cylinder (imaginary!).
  • Only one head reads/writes at any one time.

15
Disks
Track
Sector
Top viewof a platter surface
Gap
16
Disks
  • Block size is a multiple of sector size (which is
    fixed).
  • Time to access (read/write) a disk block (disk
    latency) consists of three components
  • - seek time moving arms to position disk head on
    track,
  • - rotational delay (waiting for block to rotate
    under head), and
  • transfer time (actually moving data to/from disk
    surface).
  • Seek time and rotational delay dominate.

17
Disks
Seek time
Time
3 or 5x
x
1
N
Cylinders Traveled
18
Disks
Average seek time
N
N
? ? SEEKTIME (i ? j) S
N(N-1)
j1 j?i
i1
Typical average seek time 5 ms
19
Disks
Average rotational delay
Head Here
Block I Want
Average rotational delay R 1/2
revolution Typical R 5 ms
20
Disks
Transfer time
Typical transfer rate 100 MB/sec Typical block
size 16KB Transfer time block size
transfer rate Typical transfer time 0.16 ms
21
Disks
  • Typical average disk latency is 10 ms, maximum
    latency 20 ms.
  • In 10 ms, a modern microprocessor can execute
    millions of instructions.
  • Thus, the time for a block access by far
    dominates the time typically needed for
    processing the data in memory.
  • The number of disk I/Os (block accesses) is a
    good approximation for the cost of a database
    operation.

22
Accelerating Disk Access
  • Organize data by cylinders to minimize the seek
    time and rotational delay.
  • Next block concept
  • - blocks on same track, followed by
  • - blocks on same cylinder, followed by
  • blocks on adjacent cylinder.
  • Blocks in a file are placed sequentially on disk
    (by next).
  • Disk latency can approach the transfer rate.

23
Accelerating Disk Access
  • Example
  • Assuming 10 ms average seek time, no rotational
    delay, 40 MB/s transfer rate.
  • Read a single 4 KB Block
  • Random I/O ? 10 ms
  • Sequential I/O ? 10 ms
  • Read 4 MB in 4 KB Blocks (amortized)
  • Random I/O ? 10 s
  • Sequential I/O ? 0.1 s? Speedup factor of 100

24
Accelerating Disk Access
  • Block size selection
  • Bigger blocks ? amortize I/O cost.
  • Bigger blocks ? read in more useless stuff and
    takes longer to read.
  • Good trade-off block size from 4KB to 16 KB.
  • With decreasing memory costs, blocks are becoming
    bigger!

25
Accelerating Disk Access
  • Using multiple disks
  • Replace one disk (with one independent head) by
    many disks (with many independent heads).
  • Striping a relation R divide its blocks over n
    disks in a round robin fashion.
  • Assuming that disk controller, bus and main
    memory can handle n times the transfer rate,
    striping a relation across n disks can lead to a
    speedup factor of up to n.

26
Accelerating Disk Access
  • Disk scheduling
  • For I/O requests from different processes, let
    the disk controller choose the processing order.
  • According to the elevator algorithm, the disk
    controller keeps sweeping from the innermost to
    the outermost cylinder, stopping at a cylinder
    for which there is an I/O request.
  • Can reverse sweep direction as soon as there is
    no I/O request ahead in the current direction.
  • Optimizes the throughput and average response
    time.

27
Accelerating Disk Access
  • Double buffering
  • In some scenarios, we can predict the order in
    which blocks will be requested from disk by some
    process.
  • Prefetching (double buffering) is the method of
    fetching the necessary blocks into the buffer in
    advance.
  • Requires enough buffer space.
  • Speedup factor up to n, where n is the number of
    blocks requested by a process.

28
Accelerating Disk Access
  • Single buffering
  • (1) Read B1 ? Buffer
  • (2) Process Data in Buffer
  • (3) Read B2 ? Buffer
  • (4) Process Data in Buffer
  • ...
  • Execution time n(PR)where
  • P time to process one block
  • R time to read in one block
  • n blocks read.

29
Accelerating Disk Access
  • Double buffering
  • (1) Read B1, . . ., Bn ? Buffer
  • (2) Process B1 in Buffer
  • (3) Process B2 in Buffer
  • ...
  • Execution time R nPas opposed to n(PR).
  • ? remember that R gtgt P

30
Disk Failures
  • In an intermittent failure, a read or write
    operation is unsuccessful, but succeeds with
    repeated tries.? parity checks to detect
    intermittent failures
  • Media decay is a permanent corruption of one or
    more bits which make the corresponding sector
    impossible to read / write.? stable storage to
    recover from media decay
  • A disk crash makes the entire disk permanently
    unreadable.? RAID to recover from disk crashes

31
Disk Failures
  • Checksums
  • Add n parity bits every m data bits.
  • The number of 1s among a collection of bits and
    their parity bit is always even.
  • The parity bit is the modulo-2 sum of its data
    bits.
  • m8, n1
  • Block A 011010001 (odd of 1s)
  • Block B 111011100 (even of 1s)
  • If Block A instead contains
  • Block A 011000001 (has odd of 1s)
  • error detected

32
Disk Failures
  • Checksums
  • But what if multiple bits are corrupted?
  • E.g., if Block A instead contains
  • Block A 010000001 (has even of 1s)
  • error cannot be detected
  • Probability that a single parity bit cannot
    detect a corrupt block is ½.
  • This is assuming that the probability of disk
    failures involving an odd / even number of bits
    is identical.

33
Disk Failures
  • Checksums
  • More parity bits decrease the probability of an
    undetected failure. With n m independent parity
    bits, this probability is only 1/2n .
  • E.g., we can have eight parity bits, one for the
    first bit of every byte, the second one for the
    second bit of every byte . . .
  • The chance for not detecting a disk failure is
    then only 1/256.

34
Disk Failures
  • Stable storage
  • Sectors are paired, and information X is written
    both on sectors Xl and Xr.
  • Assume that both copies are written with a
    sufficient number of parity bits so that bad
    sectors can be detected.
  • If sector is bad (according to checksum), write
    to alternative sector.
  • Alternate reading Xl and Xr until a good value is
    returned.
  • Probability of Xl and Xr both failing is very low.

35
Disk Failures
  • Disk arrays
  • So far, we cannot recover from disk crashes.
  • To address this problem, use Redundant Arrays of
    Independent Disks (RAID), arrangements of several
    disks that gives abstraction of a single, large
    disk.
  • Goals Increase reliability (and performance).
  • Redundant information allows reconstruction of
    data if a disk fails.
  • Data striping improves the disk performance.

36
Disk Failures
  • Failure Models for Disks
  • What is the expected time until disk crash?
  • We assume uniform distribution of failures over
    time.
  • Mean time to failure time period by which 50 of
    a population of disks have failed (crashed).
  • Typical mean time to failure is 10 years.
  • In this case, 5 of disks crash in the first
    year, 5 crash in the second year, . . ., 5
    crash in the tenth year, . . ., 5 crash in the
    twentieth year.

37
Disk Failures
  • Failure Models for Disks
  • Given the mean time to failure (mtf) in years, we
    can derive the probability p of a particular disk
    failing in a given year.
  • p 1 / (2 mtf)
  • Ex. mtf 10, p 1/20 5
  • Mean time to data loss time period by which 50
    of a population of disks have had a crash that
    resulted in data loss.
  • The mean time to disk failure is not necessarily
    the same as the mean time to data loss.

38
Disk Failures
  • Failure Models for Disks
  • Failure rate percentage of disks of a population
    that have failed until a certain point of time.
  • Survival rate percentage of disks of a
    population that have not failed until a certain
    point of time.
  • While it simplifies the analysis, the assumption
    of uniform distribution of failures is
    unrealistic.
  • Disks tend to fail early (manufacturing defects
    that have not been detected) or late
    (wear-and-tear).

39
Disk Failures
  • Failure Models for Disks
  • Survival rate (realistic)
  • time

40
Disk Failures
  • Mirroring
  • The data disk is copied unto a second disk, the
    mirror disk.
  • When one of the disk crashes, we replace it by a
    new disk and copy the other disk to the new one.
  • Data loss can only occur if the second disk
    crashes while the first one is being replaced.
  • This probability is negligible.
  • Mirroring is referred to as RAID level 1.

41
Disk Failures
  • Parity blocks
  • Mirroring doubles the number of disks needed.
  • The parity block approach needs only one
    redundant disk for n (arbitray) data disks.
  • In the redundant disk, the ith block stores
    parity checks for the ith blocks of all the n
    data disks.
  • Parity block approach is called RAID level 4.

A
B
C
P
42
Disk Failures
  • Parity blocks
  • Reading blocks is the same as without parity
    blocks.
  • When writing a block on a data disk, we also need
    to update the corresponding block of the
    redundant disk.
  • This can be done using four (three additional)
    disk I/O read old value of data disk block, read
    corresponding block of redundant (parity) disk,
    write new data block, recompute and write new
    redundant block.

43
Disk Failures
  • Parity blocks
  • If one of the disks crashes, we bring in a new
    disk.
  • The content of this disk can be computed, bit by
    bit, using the remaining n disks.
  • No difference between data disks and parity disk.
  • Computation based on the definition of parity,
    i.e. total number of ones is even.

44
Disk Failures
Example
  • n 3 data disks
  • Disk 1, block 1 11110000
  • Disk 2, block 1 10101010
  • Disk 3, block 1 00111000
  • and one parity disk
  • Disk 4, block 1 01100010
  • ? Sum over each column is always an even number
    of 1s
  • ? Mod-2 sum can recover any missing single row

45
Disk Failures
Example
  • Suppose we have
  • Disk 1, block 1 11110000
  • Disk 2, block 1 ????????
  • Disk 3, block 1 00111000
  • Disk 4, block 1 01100010 (parity)
  • Use mod-2 sums for block 1 over disks 1,3,4 to
    recover block 1 of failed disk 2
  • ? Disk 2, block 1 10101010

46
Disk Failures
  • RAID level 5
  • In the RAID 4 scheme, the parity disk is the
    bottleneck. On average, n-times as many writes on
    the parity disk as on the data disks.
  • However, the failure recovery method does not
    distinguish the types of the n 1 disks.
  • RAID level 5 does not use a fixed parity disk,
    but use block i of disk j as redundant if i MOD
    n1 j.

D
C
B
A
Write a Comment
User Comments (0)
About PowerShow.com