Algorithms and Methods for Distributed Storage Networks 1. Motivation, Organization, Overview Christian Schindelhauer

Transcription

1 Algorithms and Methods for 1. Motivation, Organization, Overview Institut für Informatik Wintersemester 2007/08

2 Organization Lecture Thursday 11 am - 1pm, room 101/SR /13 Friday 11 am - 12 pm, room 101/SR /13 Exercise (Stefan Rührup) starts Oct 29, 2008 Friday 12-1 pm, room 101/SR /13 appear every Friday on the web-pages solved voluntarily by students are the bases for the oral exam solutions of the exercises are discussed in the following week 2

3 Web Web page vorlesung/distributed-storage-w08 Slides, exercises, link to forum Forum for discussion, links, funnies etc. viewforum.php?f=28 Peer-to-Peer-Networks Summer Computer Networks and Telematics

4 Exam Dates by appointment possible dates - Monday, Thursday, Tuesday, Contact me during the lecture or send an to Oral exam based on the lecture and the exercises Mandatory registration Students of computer science register at the secretary of exams (Prüfungssekretariat) Peer-to-Peer-Networks Summer Computer Networks and Telematics

5 Algorithms and Methods for Motivation Evolution of Disks 5

6 Figure 5 Storage system volumetric density trend Figure 6 Cost of storage at the disk drive and system level VOLUMETRIC DENSITY, GBITS/CUBIC INCH Fall in Storage RANGE OF POSSIBLE VALUES Prices Technological impact of magnetic SYSTEMS W/LARGE SYSTEMS W/SMALL hard disk drives on storage FF DRIVES FF DRIVES systems, Grochowski, R. D. Halem IBM SYSTEMS JOURNAL, VOL 42, NO 2, PRICE/MBYTE, DOLLARS SYSTEMS W/LARGE FF DRIVES SMALL FF HDD STORAGE SYSTEM SYSTEMS W/SMALL FF DRIVES DESKTOP HDD INDUSTRY PROJECTION FROM MULTIPLE SOURCES HIGH PERFOR- MANCE, SERVER HDD Distributed age had Storage significantly Networksundercut film and paper costs by the late 1990s. It is obvious that this is a principal 6 advantage to the growth and popularity of on-line (i.e., nonarchival) magnetic storage. Concurrently, there has been an accompanying in- Figure E) Cost of storage for disk drive, paper, film, and semiconductor memory SEMICONDUCTOR MEMORY

7 age had significantly undercut film and paper costs by the late 1990s. It is obvious that this is a principal advantage Price to the growth Fall and popularity of of on-line (i.e., nonarchival) magnetic storage. RAM and Disk Concurrently, there has been an accompanying increase in the quantity of DRAM found in most storage subsystems. Storage When first introduced in the mid- 1980s, a system cache of 8 megabytes was considered large. Today, a system cache of 8 gigabytes is considered relatively small. The corresponding decrease in price-per-storage capacity of DRAM, although prices started at a higher level than magnetic hard disk drives, has enabled larger caches while still maintaining a price reduction per capacity at the system level, as Figure 6 indicates. Figure PRICE (DOLLARS PER MEGABYTE) Cost of storage for disk drive, paper, film, and semiconductor memory DISK DRIVES PAPER AND FILM SEMICONDUCTOR MEMORY (DRAM AND FLASH) 2.5 INCH Form factor miniaturization Technological impact of magnetic Figure hard 8disk displays drives form-factor on storage miniaturization systems, for disk drives Grochowski, for the time R. D. period Halem Large formfactor IBM drives, SYSTEMS withjournal, 14- and 10.8-inch VOL 42, nominal NO 2, 2003 sizes, were replaced by 3.5-inch form-factor drives at about 1995, and it is possible that in the future even smaller INCH IBM SYSTEMS JOURNAL, VOL 42, NO 2, 2003 GROCHOWSKI AND HALEM 341 7

8 Figure 1 Hard disk drive areal density trend Figure 2 Stora stora Increase of Density Technological impact of magnetic hard disk drives on storage systems, Grochowski, R. D. Halem IBM SYSTEMS JOURNAL, VOL 42, NO 2, 2003 AREAL DENSITY MEGABITS/SQUARE INCH IBM DISK DRIVE PRODUCTS INDUSTRY LAB DEMOS 100% CGR 60% CGR 25% CGR 1st AFC MEDIA 1st GMR HEAD 1st MR HEAD 1st THIN FILM HEAD IBM RAMAC (FIRST HARD DISK DRIVE) RANGE OF POSSIBLE VALUES ~35 MILLION X INCREASE RAW CAPACITY PER FLOOR SPACE AREA, GBYTES/SQUARE FOOT % C 25% C IBM RA 1960 AFC=ANTIFERROMAGNETICALLY COUPLED GMR=GIANT MAGNETORESISTIVE ESS=IBM TOT 8 observed since the first disk drive. Although the CGR is expected to eventually decline, based on increased processing difficulties in magnetic head and disk media production, laboratory demonstrations of advanced head designs indicate that a head with a capacity of greater than 100 gigabits/in 2 is feasible to store one t 45 years is sh a factor of 10 7 areal density.

9 Figure 1 AREAL DENSITY MEGABITS/SQUARE INCH Hard disk drive areal density trend 1st AFC MEDIA 10 4 (Floor 100% CGR Space) 1st GMR HEAD 10 3 hard 10 2 disk drives on storage systems, 10 3 IBM DISK DRIVE PRODUCTS Increase INDUSTRY LAB DEMOS of RANGE OF POSSIBLE VALUES 60% CGR 25% CGR Density 1st MR HEAD 1st THIN FILM HEAD Technological impact of magnetic ~35 MILLION X INCREASE Grochowski, IBM RAMAC R. D. (FIRST Halem HARD DISK DRIVE) IBM SYSTEMS JOURNAL, VOL 42, NO 2, AFC=ANTIFERROMAGNETICALLY COUPLED GMR=GIANT MAGNETORESISTIVE Figure 2 RAW CAPACITY PER FLOOR SPACE AREA, GBYTES/SQUARE FOOT Storage floor space utilization trend IBM storage systems 60% CGR 25% CGR 3380 IBM ESS MOD 800 ESS MOD F RAMAC 2 IBM RAMAC (FIRST HARD DISK DRIVE) RANGE OF POSSIBLE VALUES ESS=IBM TOTALSTORAGE ENTERPRISE STORAGE SERVER observed since the first disk drive. Although the CGR 9 is expected to eventually decline, based on increased processing difficulties in magnetic head and disk media production, laboratory demonstrations of advanced head designs indicate that a head with a capacity of greater than 100 gigabits/in 2 is feasible to store one terabyte of information over the past 45 years is shown to have decreased by more than a factor of 10 7, similar to increases in server disk drive areal density.

10 Increase of Density (Floor Space) Technological impact of magnetic hard disk drives on storage systems, Grochowski, R. D. Halem IBM SYSTEMS JOURNAL, VOL 42, NO 2, 2003 Figure FLOOR SPACE AREA, SQUARE FOOT Floor space required to store 1 terabyte IBM RAMAC (FIRST HARD DISK DRIVE) 3380 RAMAC 2 ESS MOD F ESS MOD 800 RANGE OF POSSIBLE VALUES disk drives, in thi factor, directly re tems with higher steeper growth tr Cost of storage The price trends pected to be depe extent on unit co other electronics tem, and various cost reductions f modest reduction expected. Howev drop in cost-perdrive costs as its sity increases wit percent to 100 per that price decline percent, respecti unit price 1 capacity n 10 Figure 4 INCH 2 Hard disk drive volumetric density trend 10 4 HARD DISK DRIVE RANGE OF FORM FACTOR POSSIBLE VALUES 14/10.8 INCH where r is the CGR time period n to picted in Figure 6 cent price decline sity CGR increase 15 years, unit pri that capacity has Figure 6 also show cline and indicate factor drives has

11 0.1 RANGE OF POSSIBLE VALUES unit price 1 capacity n Increase of Density (Cubic Space) Technological impact of magnetic hard disk drives on storage systems, Grochowski, R. D. Halem IBM SYSTEMS JOURNAL, VOL 42, NO 2, 2003 Figure 4 VOLUMETRIC DENSITY, GBITS/CUBIC INCH Hard disk drive volumetric density trend HARD DISK DRIVE FORM FACTOR 14/10.8 INCH RANGE OF POSSIBLE VALUES CGR = 100% CGR = 25% where r is the CGR time period n to picted in Figure 6 cent price decline sity CGR increase 15 years, unit pri that capacity has Figure 6 also show cline and indicate factor drives has price decline, ne trend for disk dr had a significant i jected that any fu increase would al storage systems. Figure 7 shows th ity decline for DR similar decline fo decrease in semic influenced the sys as can be seen in t pacity trend. How provement that i in the trends not served from Figu GROCHOWSKI AND HALEM

12 Evolution of Disk Form Factors Figure 8 The evolution of disk drive form factors INCH FORM FACTOR 5 MB (NOT SHOWN) 14-INCH FORM FACTOR 3.8 GB form-factor drives, such as 2.5-inch drives, will stitute the storage devices used for large storag rays. 4 In the past, only eight large-form-factor drives could be contained within a storage sy frame of reasonable size, whereas today s sys may contain as many as 256 drives (or more) RAID (redundant array of independent disk equivalent configuration. The availability of sm form-factor drives has been directly responsibl the development and use of system architec such as RAID, which have advanced storage to levels of low cost, capacity, performance, and ability. In addition to the increases in relia brought about by these enhanced drive technolo the adoption of RAID architectures has resulte even higher levels of improvement in system a ability. RAID architectures enable the effective coupling of hardware reliability from system a ability. 8-INCH FORM FACTOR 65 MB 10.8-INCH FORM FACTOR 5.7 GB The miniaturization trend in disk drives, sim neous with the trend of increasing drive capaci the direct result of the vast technology improvem that have increased areal density over seven or of magnitude since These improvements principally been: the introduction of magnetor tive (MR) and giant magnetoresistive (GMR) heads, finer-line-width inductive write head ments, high-signal-amplitude thin film disk ma als, lower flying heights, and partial-responseimum-likelihood (PRML) data channels. These many more innovations demonstrate the contin trend toward further enhancements. 5 Technological impact of magnetic hard disk drives on storage systems, Grochowski, R. D. Halem IBM SYSTEMS JOURNAL, VOL 42, NO 2, INCH FORM FACTOR 120 GB 3.25-INCH FORM FACTOR 84-MM DISKS/ RPM 146 GB 70-MM DISKS/ RPM 36 GB Power/performance Smaller-form-factor drives containing smaller-d eter disks with high areal density permit faster rotation rates while maintaining moderate pow quirements. Figure 9 shows this trend for disk d and storage systems, and indicates that both become significantly more economical with res to power utilization. The power loss due to air s for a disk drive is given by INCH FORM FACTOR 60 GB P constant D 4.6 R INCH FORM FACTOR 1 GB where D is disk diameter and R is the rotation Reducing disk diameter from 14 inches (355 to 65 mm would allow the drive designer to incr rotation rate from 3600RPM (revolution per min to RPM in today s higher performance drives, while maintaining acceptable power lo 342 GROCHOWSKI AND HALEM IBM SYSTEMS JOURNAL, VOL 42, NO 2

13 IBM SYSTEMS JOURNAL, VOL 42, NO 2, 2003 GROCHOWSKI AND HALEM drive design and operation, primarily the adoption of MR heads that allowed higher linear density, smaller-form-factor drives, and higher disk rotation rates. Note that at a little over 100 megabytes/s internal clock operation within a data channel circuit is approaching 1 GHz as fast as today s microprocessors and also approaching the limit for silicon circuits. Increase of Speed Technological impact of magnetic hard disk drives on storage systems, Grochowski, R. D. Halem IBM SYSTEMS JOURNAL, VOL 42, NO 2, 2003 Figure 11 shows the decrease in average seek and accessing times for server-platform disk drives over the past 30 years. Access time is defined as seek time plus latency time, which is inversely proportional to rotation rate. Seek time depends on data band, the difference between outer-disk recording radius and inner-disk recording radius (the region where data are actually recorded). The use of smaller-diameter drives causes a marked change in slope near 1991, similar to trends shown in Figure The smaller drives are faster, store more information, and consume less power per capacity. Rotation rate continues to be a key parameter in disk drive and system performance, a fact that sometimes is not apparent. Regardless of the inclusion of large DRAM caches, when the server system requests a read of data not contained within the cache, latency becomes the key delaying element. It is the application of small-formfactor drives that has allowed rotation rates to go beyond RPM and to continually improve sys- tem performance while maintaining reasonable power requirements. Figure 10 MAXIMUM INTERNAL DATA RATE, MBYTES/SECONDS Hard disk drive maximum internal data rate for enterprise/server drives 40% CGR <10% CGR DATA RATE = LINEAR x RPM x DISK DENSITY DIAMETER AVAILABILITY YEAR What impact does drive design have on system performance? Initially, system performance was essentially drive performance. After the introduction of electronic caches, system performance was related to system 13 channel speed, and cache effectiveness became the critical issue. This is shown by

14 Increase of Speed Figure 11 Disk drive access/seek times 100 TIME, MILLISECONDS 10 ULTRASTAR XP ULTRASTAR 2XP ULTRASTAR 18XP ULTRASTAR 9ZX ULTRASTAR 36XP ULTRASTAR 18ZX ULTRASTAR 36ZX ULTRASTAR 73LZX ULTRASTAR 36LZX ULTRASTAR 36Z15 SEEK TIME ~ (ACTUATOR INERTIALPOWER) 1/3 x (DATA BAND) 2/3 ROTATIONAL TIME (LATENCY) ~ (RPM) -1 ACCESS TIME = SEEK TIME + LATENCY AVAILABILITY YEAR SEEKING ACCESSING Technological impact of magnetic hard disk drives Figure on storage 12 systems, storage systems Grochowski, R. D. Halem IBM SYSTEMS JOURNAL, VOL 42, NO 2, 2003 HARD DISK DRIVE STORAGE 1000 MODELED DATA SYSTEMS I/O S PER SECOND Performance trend for hard disk drives and BASIS: 4K RECORDS ADDITION OF SYSTEM CACHE 14 System performance is a weighted average of the two performances based on the cache hit ratio. Nearly concurrent with this trend, a disk drive buffer was also included, which enhanced the performance of this component significantly. Figure 12 shows Rechnernetze relative performance of both drivesalbert-ludwigs-universität and systems over Freiburg und Telematik time, and the effect of caching and buffering is clearly visible. The basis for the chart is the storing of 4K records. Although system (and drive) performance would be expected to increase as a result of component technology advances and drive miniaturization, enhancements in electronics have been a sig-

15 Algorithms and Methods for Motivation Consumer Behavior 15

16 Consumer Usage Consumer Survey on Digital Storage in Consumer Electronics 2008, Coughlin Associates (Dec. 2007) 51% said that 1 TB disk would be useful Most storage of content was on hard disk 46% backup data less than once per year - except pictures most of them do not backup - but most think it is important to have backups out of their homes Most people want to store entire TV series, copies of their entire music collection Projection by 2013 average home has 9 Terabyte by 2015 user content sums up to 650 Exabyte 16

17 Storage Hierarchy Primary storage Processors registers Processor cache RAM Secondary storage Hard disks Solid state disks CD, DVD Tertiary storage tape libraries optical jukeboxes 17

18 Characteristics of Storage Volatile non-volatile memory non-volatile: dynamic or static Read & write Read only Slow write, fast read Random access Sequential access Addressability location addressable file addressable content addressable Capacity Performance Latency Throughput 18

19 Non-volatile Storage Technologies Punch cards (Hollerith) s Magnetic tape data storage 1951-today Hard disk drive 1956-today Floppy disks 1970s-1990s EEPROM (Electrically Erasable Programmable Read-Only Memory) 1980-today Flash memory Optical disc drive (read/write) 1997-today 19

20 Network Storage Types Direct attached storage (DAS) traditional storage Network attached storage (NAS) storage attached to another computer accessible at file level over LAN or WAN Storage area network (SAN) specialized network providing other computers with storage capacity with access on block-addressing level File area network (FAN) systematic approach to organize file-related storage systems organization wide high-level storage network 20

21 Overview Basic Storage Technology Hard disks Flash memory, solid state disks Storage device design File systems Classic file systems Network and distributed file systems Storage organization SAN, NAS, FAN Storage hierarchies, Tiers Selected topics of Distributed Algorithms Conflict resolution Cache strategies Redundancy RAID levels Coding techniques Internet and storage TCP/IP, FTP, Webdav, etc. Distributed Storage Systems Online storage - e.g. Amazon S3, Google Shared Storage Peer-to-peer network storage - e.g. Oceanstore 21

22 Algorithms and Methods for Distributed Storage Networks 1. Organization & Overview Institut für Informatik Wintersemester 2007/08