1. CS252/Kubiatowicz Lec 22.1 11/16/03 CS252 Graduate Computer Architecture Lecture 22 I/O Introduction November 16 th, 2003 Prof. John Kubiatowicz http://www.cs.berkeley.edu/~kubitron/courses/cs252-F03

2. CS252/Kubiatowicz Lec 22.2 11/16/03 Motivation: –DRAM is dense  Signals are easily disturbed –High Capacity  higher probability of failure Approach: Redundancy –Add extra information so that we can recover from errors –Can we do better than just create complete copies? Block Codes: Data Coded in blocks –k data bits coded into n encoded bits –Measure of overhead: Rate of Code: K/N –Often called an (n,k) code –Consider data as vectors in GF(2) [ i.e. vectors of bits ] Code Space is set of all 2 n vectors, Data space set of 2 k vectors –Encoding function: C=f(d) –Decoding function: d=f(C’) –Not all possible code vectors, C, are valid! Systematic Codes: original data appears within coded data Review: Error Correction

3. CS252/Kubiatowicz Lec 22.3 11/16/03 Not every vector in the code space is valid Hamming Distance (d): –Minimum number of bit flips to turn one code word into another Number of errors that we can detect: (d-1) Number of errors that we can fix: ½(d-1) Code Space d0d0 C 0 =f(d 0 ) Code Distance (Hamming Distance) General Idea: Code Vector Space

4. CS252/Kubiatowicz Lec 22.4 11/16/03 Review: Code Types Linear Codes: Code is generated by G and in null-space of H Hamming Codes: Design the H matrix –d = 3  Columns nonzero, Distinct –d = 4  Columns nonzero, Distinct, Odd-weight Reed-solomon codes: –Based on polynomials in GF(2 k ) (I.e. k-bit symbols) –Data as coefficients, code space as values of polynomial: –P(x)=a 0 +a 1 x 1 +… a k-1 x k-1 –Coded: P(0),P(1),P(2)….,P(n-1) –Can recover polynomial as long as get any k of n –Alternatively: as long as no more than n-k coded symbols erased, can recover data. Side note: Multiplication by constant in GF(2 k ) can be represented by k  k matrix: a  x –Decompose unknown vector into k bits: x=x 0 +2x 1 +…+2 k-1 x k-1 –Each column is result of multiplying a by 2 i

5. CS252/Kubiatowicz Lec 22.5 11/16/03 Motivation: Who Cares About I/O? CPU Performance: 60% per year I/O system performance limited by mechanical delays (disk I/O) < 10% per year (IO per sec or MB per sec) Amdahl's Law: system speed-up limited by the slowest part! 10% IO & 10x CPU => 5x Performance (lose 50%) 10% IO & 100x CPU => 10x Performance (lose 90%) I/O bottleneck: Diminishing fraction of time in CPU Diminishing value of faster CPUs

6. CS252/Kubiatowicz Lec 22.6 11/16/03 I/O Systems Processor Cache Memory - I/O Bus Main Memory I/O Controller Disk I/O Controller I/O Controller Graphics Network interrupts

7. CS252/Kubiatowicz Lec 22.7 11/16/03 Technology Trends Disk Capacity now doubles every 18 months; before 1990 every 36 motnhs Today: Processing Power Doubles Every 18 months Today: Memory Size Doubles Every 18 months(4X/3yr) Today: Disk Capacity Doubles Every 18 months Disk Positioning Rate (Seek + Rotate) Doubles Every Ten Years! The I/O GAP The I/O GAP

8. CS252/Kubiatowicz Lec 22.8 11/16/03 Storage Technology Drivers Driven by the prevailing computing paradigm –1950s: migration from batch to on-line processing –1990s: migration to ubiquitous computing »computers in phones, books, cars, video cameras, … »nationwide fiber optical network with wireless tails Effects on storage industry: –Embedded storage »smaller, cheaper, more reliable, lower power –Data utilities »high capacity, hierarchically managed storage

9. CS252/Kubiatowicz Lec 22.9 11/16/03 Historical Perspective 1956 IBM Ramac — early 1970s Winchester –Developed for mainframe computers, proprietary interfaces –Steady shrink in form factor: 27 in. to 14 in. 1970s developments –5.25 inch floppy disk formfactor (microcode into mainframe) –early emergence of industry standard disk interfaces »ST506, SASI, SMD, ESDI Early 1980s –PCs and first generation workstations Mid 1980s –Client/server computing –Centralized storage on file server »accelerates disk downsizing: 8 inch to 5.25 inch –Mass market disk drives become a reality »industry standards: SCSI, IPI, IDE »5.25 inch drives for standalone PCs, End of proprietary interfaces

10. CS252/Kubiatowicz Lec 22.10 11/16/03 Disk History Data density Mbit/sq. in. Capacity of Unit Shown Megabytes 1973: 1. 7 Mbit/sq. in 140 MBytes 1979: 7. 7 Mbit/sq. in 2,300 MBytes source: New York Times, 2/23/98, page C3, “Makers of disk drives crowd even mroe data into even smaller spaces”

11. CS252/Kubiatowicz Lec 22.11 11/16/03 Historical Perspective Late 1980s/Early 1990s: –Laptops, notebooks, (palmtops) –3.5 inch, 2.5 inch, (1.8 inch formfactors) –Formfactor plus capacity drives market, not so much performance »Recently Bandwidth improving at 40%/ year –Challenged by DRAM, flash RAM in PCMCIA cards »still expensive, Intel promises but doesn’t deliver »unattractive MBytes per cubic inch –Optical disk fails on performace (e.g., NEXT) but finds niche (CD ROM)

12. CS252/Kubiatowicz Lec 22.12 11/16/03 Disk History 1989: 63 Mbit/sq. in 60,000 MBytes 1997: 1450 Mbit/sq. in 2300 MBytes source: New York Times, 2/23/98, page C3, “Makers of disk drives crowd even mroe data into even smaller spaces” 1997: 3090 Mbit/sq. in 8100 MBytes

13. CS252/Kubiatowicz Lec 22.13 11/16/03 MBits per square inch: DRAM as % of Disk over time source: New York Times, 2/23/98, page C3, “Makers of disk drives crowd even mroe data into even smaller spaces” 470 v. 3000 Mb/si 9 v. 22 Mb/si 0.2 v. 1.7 Mb/si

14. CS252/Kubiatowicz Lec 22.14 11/16/03 Disk Performance Model /Trends Capacity + 100%/year (2X / 1.0 yrs) Transfer rate (BW) + 40%/year (2X / 2.0 yrs) Rotation + Seek time – 8%/ year (1/2 in 10 yrs) MB/$ > 100%/year (2X / <1.5 yrs) Fewer chips + areal density

15. CS252/Kubiatowicz Lec 22.15 11/16/03 Photo of Disk Head, Arm, Actuator Actuator Arm Head Platters (12) { Spindle

16. CS252/Kubiatowicz Lec 22.16 11/16/03 Nano-layered Disk Heads Special sensitivity of Disk head comes from “Giant Magneto- Resistive effect” or (GMR) IBM is leader in this technology –Same technology as TMJ-RAM breakthrough we described in earlier class. Coil for writing

17. CS252/Kubiatowicz Lec 22.17 11/16/03 Disk Device Terminology Several platters, with information recorded magnetically on both surfaces (usually) Actuator moves head (end of arm,1/surface) over track (“seek”), select surface, wait for sector rotate under head, then read or write – “Cylinder”: all tracks under heads Bits recorded in tracks, which in turn divided into sectors (e.g., 512 Bytes) Platter Outer Track Inner Track Sector Actuator HeadArm

18. CS252/Kubiatowicz Lec 22.18 11/16/03 Disk Performance Example Disk Latency = Queuing Time + Seek Time + Rotation Time + Xfer Time + Ctrl Time Order of magnitude times for 4K byte transfers: Seek: 12 ms or less Rotate: 4.2 ms @ 7200 rpm = 0.5 rev/(7200 rpm/60m/s) (8.3 ms @ 3600 rpm ) Xfer: 1 ms @ 7200 rpm (2 ms @ 3600 rpm) Ctrl: 2 ms (big variation) Disk Latency = Queuing Time + (12 + 4.2 + 1 + 2)ms = QT + 19.2ms Average Service Time = 19.2 ms

19. CS252/Kubiatowicz Lec 22.19 11/16/03 Disk Time Example Disk Parameters: –Transfer size is 8K bytes –Advertised average seek is 12 ms –Disk spins at 7200 RPM –Transfer rate is 4 MB/sec Controller overhead is 2 ms Assume that disk is idle so no queuing delay What is Average Disk Access Time for a Sector? –Ave seek + ave rot delay + transfer time + controller overhead –12 ms + 0.5/(7200 RPM/60) + 8 KB/4 MB/s + 2 ms –12 + 4.15 + 2 + 2 = 20 ms Advertised seek time assumes no locality: typically 1/4 to 1/3 advertised seek time: 20 ms => 12 ms

20. CS252/Kubiatowicz Lec 22.20 11/16/03 State of the Art: Ultrastar 72ZX –73.4 GB, 3.5 inch disk –2¢/MB –10,000 RPM; 3 ms = 1/2 rotation –11 platters, 22 surfaces –15,110 cylinders –7 Gbit/sq. in. areal den –17 watts (idle) –0.1 ms controller time –5.3 ms avg. seek –50 to 29 MB/s(internal) source: www.ibm.com; www.pricewatch.com; 2/14/00 Latency = Queuing Time + Controller time + Seek Time + Rotation Time + Size / Bandwidth per access per byte { + Sector Track Cylinder Head Platter Arm Track Buffer

21. CS252/Kubiatowicz Lec 22.21 11/16/03 CS 252 Administrivia Upcoming schedule of project events in CS 252 –People who have not talked to me about their projects should do so soon! –Oral Presentations: Tue/Wed 12/9-10 –Final projects due: Friday 12/12? Final topics: –Some queuing theory –Multiprocessing Final Midterm: 12/1 (Monday after thanksgiving) Alternative: 12/3?

22. CS252/Kubiatowicz Lec 22.22 11/16/03 Alternative Data Storage Technologies: Early 1990s CapBPITPIBPI*TPIData Xfer Access Technology(MB)(Million)(KByte/s) Time Conventional Tape: Cartridge (.25")15012000104 1.2 92minutes IBM 3490 (.5")8002286038 0.93000seconds Helical Scan Tape: Video (8mm)4600432001638 71 49245 secs DAT (4mm)1300610001870114 18320 secs Magnetic & Optical Disk: Hard Disk (5.25") 1200 335281880 63300018 ms IBM 3390 (10.5") 3800 279402235 62425020 ms Sony MO (5.25") 6402413018796454 88100 ms

23. CS252/Kubiatowicz Lec 22.23 11/16/03 Tape vs. Disk Longitudinal tape uses same technology as hard disk; tracks its density improvements Disk head flies above surface, tape head lies on surface Disk fixed, tape removable Inherent cost-performance based on geometries: fixed rotating platters with gaps (random access, limited area, 1 media / reader) vs. removable long strips wound on spool (sequential access, "unlimited" length, multiple / reader) New technology trend: Helical Scan (VCR, Camcoder, DAT) Spins head at angle to tape to improve density

24. CS252/Kubiatowicz Lec 22.24 11/16/03 Current Drawbacks to Tape Tape wear out: –Helical 100s of passes to 1000s for longitudinal Head wear out: –2000 hours for helical Both must be accounted for in economic / reliability model Long rewind, eject, load, spin-up times; not inherent, just no need in marketplace (so far) Designed for archival

25. CS252/Kubiatowicz Lec 22.25 11/16/03 Automated Cartridge System STC 4400 6000 x 0.8 GB 3490 tapes = 5 TBytes in 1992 $500,000 O.E.M. Price 6000 x 10 GB D3 tapes = 60 TBytes in 1998 Library of Congress: all information in the world; in 1992, ASCII of all books = 30 TB 8 feet 10 feet

26. CS252/Kubiatowicz Lec 22.26 11/16/03 Relative Cost of Storage Technology—Late 1995/Early 1996 Magnetic Disks 5.25”9.1 GB$2129$0.23/MB $1985$0.22/MB 3.5”4.3 GB$1199$0.27/MB $999$0.23/MB 2.5”514 MB$299$0.58/MB 1.1 GB$345$0.33/MB Optical Disks 5.25”4.6 GB$1695+199$0.41/MB $1499+189$0.39/MB PCMCIA Cards Static RAM4.0 MB$700$175/MB Flash RAM40.0 MB$1300 $32/MB 175 MB$3600$20.50/MB

27. CS252/Kubiatowicz Lec 22.27 11/16/03 Manufacturing Advantages of Disk Arrays 14” 10”5.25”3.5” Disk Array: 1 disk design Conventional: 4 disk designs Low End High End Disk Product Families

28. CS252/Kubiatowicz Lec 22.28 11/16/03 Replace Small # of Large Disks with Large # of Small Disks! (1988 Disks) Data Capacity Volume Power Data Rate I/O Rate MTTF Cost IBM 3390 (K) 20 GBytes 97 cu. ft. 3 KW 15 MB/s 600 I/Os/s 250 KHrs $250K IBM 3.5" 0061 320 MBytes 0.1 cu. ft. 11 W 1.5 MB/s 55 I/Os/s 50 KHrs $2K x70 23 GBytes 11 cu. ft. 1 KW 120 MB/s 3900 IOs/s ??? Hrs $150K Disk Arrays have potential for large data and I/O rates high MB per cu. ft., high MB per KW reliability?

29. CS252/Kubiatowicz Lec 22.29 11/16/03 Array Reliability Reliability of N disks = Reliability of 1 Disk ÷ N 50,000 Hours ÷ 70 disks = 700 hours Disk system MTTF: Drops from 6 years to 1 month! Arrays (without redundancy) too unreliable to be useful! Hot spares support reconstruction in parallel with access: very high media availability can be achieved Hot spares support reconstruction in parallel with access: very high media availability can be achieved

30. CS252/Kubiatowicz Lec 22.30 11/16/03 Redundant Arrays of Disks Files are "striped" across multiple spindles Redundancy yields high data availability Disks will fail Contents reconstructed from data redundantly stored in the array Capacity penalty to store it Bandwidth penalty to update Mirroring/Shadowing (high capacity cost) Horizontal Hamming Codes (overkill) Parity & Reed-Solomon Codes Failure Prediction (no capacity overhead!) VaxSimPlus — Technique is controversial Techniques:

31. CS252/Kubiatowicz Lec 22.31 11/16/03 Redundant Arrays of Disks RAID 1: Disk Mirroring/Shadowing Each disk is fully duplicated onto its "shadow" Very high availability can be achieved Bandwidth sacrifice on write: Logical write = two physical writes Reads may be optimized Most expensive solution: 100% capacity overhead Targeted for high I/O rate, high availability environments recovery group

32. CS252/Kubiatowicz Lec 22.32 11/16/03 Redundant Arrays of Disks RAID 3: Parity Disk P 10010011 11001101 10010011... logical record 1001001110010011 1100110111001101 1001001110010011 0011000000110000 Striped physical records Parity computed across recovery group to protect against hard disk failures 33% capacity cost for parity in this configuration wider arrays reduce capacity costs, decrease expected availability, increase reconstruction time Arms logically synchronized, spindles rotationally synchronized logically a single high capacity, high transfer rate disk Targeted for high bandwidth applications: Scientific, Image Processing

33. CS252/Kubiatowicz Lec 22.33 11/16/03 Redundant Arrays of Disks RAID 5+: High I/O Rate Parity A logical write becomes four physical I/Os Independent writes possible because of interleaved parity Reed-Solomon Codes ("Q") for protection during reconstruction A logical write becomes four physical I/Os Independent writes possible because of interleaved parity Reed-Solomon Codes ("Q") for protection during reconstruction D0D1D2 D3 P D4D5D6 P D7 D8D9P D10 D11 D12PD13 D14 D15 PD16D17 D18 D19 D20D21D22 D23 P.............................. Disk Columns Increasing Logical Disk Addresses Stripe Unit Targeted for mixed applications

34. CS252/Kubiatowicz Lec 22.34 11/16/03 Problems of Disk Arrays: Small Writes D0D1D2 D3 P D0' + + D1D2 D3 P' new data old data old parity XOR (1. Read) (2. Read) (3. Write) (4. Write) RAID-5: Small Write Algorithm 1 Logical Write = 2 Physical Reads + 2 Physical Writes

35. CS252/Kubiatowicz Lec 22.35 11/16/03 Subsystem Organization host array controller single board disk controller single board disk controller single board disk controller single board disk controller host adapter manages interface to host, DMA control, buffering, parity logic physical device control often piggy-backed in small format devices striping software off-loaded from host to array controller no applications modifications no reduction of host performance

36. CS252/Kubiatowicz Lec 22.36 11/16/03 System Availability: Orthogonal RAIDs Array Controller String Controller String Controller String Controller String Controller String Controller String Controller... Data Recovery Group: unit of data redundancy Redundant Support Components: fans, power supplies, controller, cables End to End Data Integrity: internal parity protected data paths

37. CS252/Kubiatowicz Lec 22.37 11/16/03 System-Level Availability Fully dual redundant I/O Controller Array Controller......... Recovery Group Goal: No Single Points of Failure Goal: No Single Points of Failure host with duplicated paths, higher performance can be obtained when there are no failures

38. CS252/Kubiatowicz Lec 22.38 11/16/03 OceanStore: Global Scale Persistent Storage Global-Scale Persistent Storage

39. CS252/Kubiatowicz Lec 22.39 11/16/03 Pac Bell Sprint IBM AT&T Canadian OceanStore Service provided by confederation of companies –Monthly fee paid to one service provider –Companies buy and sell capacity from each other IBM Utility-based Infrastructure

40. CS252/Kubiatowicz Lec 22.40 11/16/03 Important P2P Technology (Decentralized Object Location and Routing) GUID1 DOLR GUID1 GUID2

41. CS252/Kubiatowicz Lec 22.41 11/16/03 The Path of an OceanStore Update Second-Tier Caches Multicast trees Inner-Ring Servers Clients

42. CS252/Kubiatowicz Lec 22.42 11/16/03 Archival Dissemination of Fragments

43. CS252/Kubiatowicz Lec 22.43 11/16/03 Aside: Why erasure coding? High Durability/overhead ratio! Exploit law of large numbers for durability! 6 month repair, FBLPY: –Replication: 0.03 –Fragmentation: 10 -35 Fraction Blocks Lost Per Year (FBLPY)

44. CS252/Kubiatowicz Lec 22.44 11/16/03 The Dissemination Process: Achieving Failure Independence Model Builder Set Creator Introspection Human Input Network Monitoring model Inner Ring set probe type fragments Anti-correlation Analysis/Models Mutual Information Human Input Data Mining Independent Set Generation Effective Dissemination

45. CS252/Kubiatowicz Lec 22.45 11/16/03 OceanStore Concepts Applied to Tape-less backup –Self-Replicating, Self-Repairing, Self-Managing –No need for actual Tape in system »(Although could be there to keep with tradition) The Berkeley PetaByte Archival Service

46. CS252/Kubiatowicz Lec 22.46 11/16/03 But: What about queue time? Or: why nonlinear response Response time = Queue + Device Service time 100% Response Time (ms) Throughput (Utilization) (% total BW) 0 100 200 300 0% Proc Queue IOCDevice Metrics: Response Time Throughput latency goes as T ser ×u/(1-u) u = utilization

47. CS252/Kubiatowicz Lec 22.47 11/16/03 Queueing Theory applies to long term, steady state behavior  Arrival rate = Departure rate Little’s Law: Mean number tasks in system = arrival rate x mean reponse time –Observed by many, Little was first to prove –Simple interpretation: you should see the same number of tasks in queue when entering as when leaving. Applies to any system in equilibrium, as long as nothing in black box is creating or destroying tasks “Black Box” Queueing System ArrivalsDepartures Introduction to Queueing Theory

48. CS252/Kubiatowicz Lec 22.48 11/16/03 Server spends a variable amount of time with customers –Weighted mean m1 = (f1 x T1 + f2 x T2 +...+ fn x Tn)/F =  p(T)xT –  2 = (f1 x T1 2 + f2 x T2 2 +...+ fn x Tn 2 )/F – m1 2 =  p(T)xT 2 - m1 2 –Squared coefficient of variance: C =  2 /m1 2 »Unitless measure (100 ms 2 vs. 0.1 s 2 ) Exponential distribution C = 1 : most short relative to average, few others long; 90% 90% 1 : further from average C=2.0 => 90% < 2.8 x average, 69% < average Avg. A Little Queuing Theory: Use of random distributions Avg. 0 ProcIOCDevice Queue server System

49. CS252/Kubiatowicz Lec 22.49 11/16/03 Disk response times C  1.5 (majority seeks < average) Yet usually pick C = 1.0 for simplicity –Memoryless, exponential dist –Many complex systems well described by memoryless distribution! Another useful value is average time must wait for server to complete current task: m1(z) –Called “Average Residual Wait Time” –Not just 1/2 x m1 because doesn’t capture variance –Can derive m1(z) = 1/2 x m1 x (1 + C) –No variance  C= 0 => m1(z) = 1/2 x m1 –Exponential  C= 1 => m1(z) = m1 A Little Queuing Theory: Variable Service Time ProcIOCDevice Queue server System Avg. 0 Time

50. CS252/Kubiatowicz Lec 22.50 11/16/03 Calculating average wait time in queue T q : –All customers in line must complete; avg time: m1  T ser = 1/  –If something at server, it takes to complete on average m1(z) »Chance server is busy = u= /  ; average delay is u x m1(z) T q = u x m1(z) + L q x T ser T q = u x m1(z) + x T q x T ser T q = u x m1(z) + u x T q T q x (1 – u) = m1(z) x u T q = m1(z) x u/(1-u) = T ser x {1/2 x (1+C)} x u/(1 – u)) Notation: average number of arriving customers/second T ser average time to service a customer userver utilization (0..1): u = x T ser T q average time/customer in queue L q average length of queue:L q = x T q m1(z) average residual wait time = T ser x {1/2 x (1+C)} A Little Queuing Theory: Average Wait Time Little’s Law Defn of utilization (u)

51. CS252/Kubiatowicz Lec 22.51 11/16/03 A Little Queuing Theory: M/G/1 and M/M/1 Assumptions so far: –System in equilibrium –Time between two successive arrivals in line are random –Server can start on next customer immediately after prior finishes –No limit to the queue: works First-In-First-Out –Afterward, all customers in line must complete; each avg T ser Described “memoryless” or Markovian request arrival (M for C=1 exponentially random), General service distribution (no restrictions), 1 server: M/G/1 queue When Service times have C = 1, M/M/1 queue T q = T ser x u x(1 + C) /(2 x (1 – u)) = T ser x u /(1 – u) T ser average time to service a customer userver utilization (0..1): u = r x T ser T q average time/customer in queue

52. CS252/Kubiatowicz Lec 22.52 11/16/03 A Little Queuing Theory: An Example processor sends 10 x 8KB disk I/Os per second, requests & service exponentially distrib., avg. disk service = 20 ms On average, how utilized is the disk? –What is the number of requests in the queue? –What is the average time spent in the queue? –What is the average response time for a disk request? Notation: average number of arriving customers/second = 10 T ser average time to service a customer = 20 ms (0.02s) userver utilization (0..1): u = x T ser = 10/s x.02s = 0.2 T q average time/customer in queue = T ser x u / (1 – u) = 20 x 0.2/(1-0.2) = 20 x 0.25 = 5 ms (0.005s) T sys average time/customer in system: T sys =T q +T ser = 25 ms L q average length of queue:L q = x T q = 10/s x.005s = 0.05 requests in queue L sys average # tasks in system: L sys = x T sys = 10/s x.025s = 0.25

53. CS252/Kubiatowicz Lec 22.53 11/16/03 A Little Queuing Theory: Yet Another Example Processor access memory over “network” DRAM service properties –speed = 9cycles+ 2cycles/word –With 8 word cachelines: T ser = 25cycles,  =1/25=.04 ops/cycle –Deterministic Servicetime! (C=0) Processor behavior –CPI=1, 40% memory, 7.5% cache misses –Rate: = 1 inst/cycle*.4*.075 =.03 ops/cycle Notation: average number of arriving customers/cycle=.03 T ser average time to service a customer= 25 cycles userver utilization (0..1): u = x T ser =.03  25=.75 T q average time/customer in queue = T ser x u x (1 + C) /(2 x (1 – u)) = (25 x 0.75 x ½)/(1 – 0.75) = 37.5 cycles T sys average time/customer in system: T sys = T q +T ser = 62.5 cycles L q average length of queue:L q = x T q =.03 x 37.5 = 1.125 requests in queue L sys average # tasks in system :L sys = x T sys =.03 x 62.5 = 1.875

54. CS252/Kubiatowicz Lec 22.54 11/16/03 Summary Disk industry growing rapidly, improves: –bandwidth 40%/yr, –areal density 60%/year, $/MB faster? queue + controller + seek + rotate + transfer Advertised average seek time benchmark much greater than average seek time in practice Response time vs. Bandwidth tradeoffs Queueing theory: or (c=1): Value of faster response time: –0.7sec off response saves 4.9 sec and 2.0 sec (70%) total time per transaction => greater productivity –everyone gets more done with faster response, but novice with fast response = expert with slow