CH 11. MASS STORAGE SYSTEMS. Typical Magnetic Disk. Typical Magnetic Disk. Mass Storage. Magnetic Disk Performance (1)

Transcription

1 CH 11. MASS STORAGE SYSTEMS adapted from textbook slides Figure 11.1 Moving-head disk mechanism Typical Magnetic Disk Typical Magnetic Disk disk arms platter track disk arm block/sector block/sector read/write heads read/write head Mass Storage Magnetic Disk Performance (1) Magnetic disks: most commonly used secondary storage RPM: 4500, 5400, 7200, when data is requested seek time: time to (mechanically) position disk arm to desired cylinder rotational latency: time to rotate until the desired sector is under the read/write disk head transfer rate: time to actually transfer sector between disk and computer (memory) Solid state disks Disks attached via I/O bus IDE, EIDE, ATA, SATA, USB, FIreWire, SCSI, transfer rate: 1 Gbits/sec seek time: 3 ~ 12ms, typically 8-9ms rotational latency: 1 / (RPM / 60) for example, for full rotation on 7200 RPM disk drive latency = 1 / (7200 / 60) = 1 / 120 = seconds = ms average latency = ½ * rotational latency (why?) average access time = average seek + average latency 1

2 Magnetic Disk Latency Magnetic Disk Performance (2) RPM Full rotation latency Average latency average I/O time = average access (seek + latency)+ (transfer amount / transfer rate) + controller overhead Example: transfer 4KB block on a 7200 RPM disk w/ 5ms average seek time, 1 Gb/sec transfer rate, 0.1ms overhead average I/O time = Magnetic Disk Performance (3) Magnetic Disk Performance (4) average I/O time = average access (seek + latency)+ (transfer amount / transfer rate) + controller overhead Example: transfer 4KB block on a 7200 RPM disk w/ 5ms average seek time, 1 Gb/sec transfer rate, 0.1ms overhead average I/O time = 5ms ms + (4KB / 1Gb/sec) + 0.1ms = 9.27ms + (4KB / 1Gb / sec) = 9.27ms ms = 9.301ms Faster RPM ==> Lower Latency Disadvantages of faster spin speed? vibration heat noise Solid State Disk Performance DVD Performance (wikipedia) transfer rate: limited by the host interface read speed: 200~500 MB/s write speed: 200~500 MB/s seek time: 0.1 ms latency: negligent no (mechanical) moving parts during operation advantages: allows random access to any location, light weight, more reliable, reduced power consumption, less heat, quiet disadvantages: cost 2

3 Blue-Ray Performance (wikipedia) Disk Drive Structure Addressed as one-dimensional arrays of logical blocks One logical block is the smallest unit of transfer Logical blocks are mapped into the sectors of the disk sequentially sector 0: the 1 st sector of the 1 st track on outer-most cylinder mapping continues through that track, then the rest of the tracks in that cylinder, then through the rest of the cylinders from outer-most to innermost 11.3 Disk Scheduling Algorithms Goal: minimize disk seek time by minimizing distance disk head has to travel between seeks OS maintains of queue of I/O requests for each device Disk controller may maintain request queue in disk buffer Given: queue of disk I/O requests (of cylinder #s) 98, 183, 37, 122, 14, 124, 65, 67 current location of disk (read/write) head head pointer: 53 Disk Scheduling: FCFS FCFS: First Come First Served requests are handled in the order they are made from current head position, go to the next request position advantage: simple, no algorithm really needed disadvantage: potentially head zigzags Disk Scheduling: FCFS Disk Scheduling: SSTF SSTF: Shortest Seek Time First from current head position, select the request with the shortest distance advantage: small total head movement distance disadvantage: starvation (similar to SJNF CPU scheduling algorithm) Figure 11.4 FCFS disk scheduling 3

4 Disk Scheduling: SSTF Disk Scheduling: SCAN also known as: Elevator algorithm starts from current head position and moves toward one end servicing all requests until it reaches one end; reverse direction and continue advantage: no zigzag disadvantage: long wait time for requests at the other end Figure 11.5 SSTF disk scheduling Disk Scheduling: SCAN Disk Scheduling: LOOK Figure 11.6 SCAN disk scheduling Disk Scheduling: C-SCAN C-SCAN: Circular SCAN scan and service in one direction only starts from one end and moves toward the other end servicing all requests until it reaches the other end; goes back to beginning without servicing; repeat more uniform wait time for all requests Disk Scheduling: C-SCAN Figure 11.7 C-SCAN disk scheduling 4

5 Disk Scheduling: C-LOOK Disk Scheduling Algorithms performance affected by file allocation method contiguous allocation: limited head movement linked or indexed allocation: greater head movement performance also affected by location of directory content and index blocks should be a separate module (not integrated into kernel) 11.5 Disk Management Physical (Low-Level) Format formatting boot block and booting bad blocks create tracks and divide into sectors each sector has: header, data area, trailer header: sector identification information trailer: ECC (error correcting code) user accessible data area: 256, 512, 1024,.., 4096 Byte-size actual sector size > published sector size data size implication? ECC a # is calculated based on data and sector sector read, same calculation is done then compared with the saved #; if not equal, sector (and data) may be bad Logical (High-Level) Format Boot Block create partitions: each partition treated as a logical disk OS dependent: puts its own data structure logical formatting: make a file system (volume) in a partition typically done in factory (Windows/Mac-formatted) header: sector identification information trailer: ECC (error correcting code) user accessible data area: 256, 512, 1024,.., 4096 Byte-size actual sector size > published sector size clusters: consist of several contiguous sectors bootstrap: initial boot-up initializes all aspects of system: CPU registers, device controllers, cache, main memory ==> BIOS (PCs) locates, loads, and runs operating system generally stored in ROM (Read-Only-Memory) convenient (no need to worry about virus, changing location) cannot be modified without changing ROM chip (mask ROM) can be erased by removing from circuit board (EPROM) can be modified on the board (EEPROM) 5

6 MS DOS Disk Layout Windows Disk Layout (Master Boot Record) boot sector Figure 11.9 Booting from disk in Windows CS Lab Machine Partition Table Bad Block: defective sector Disk /dev/sda: GB, bytes 63 sectors/track, cylinders Sector size (logical/physical): 512 bytes / 512 bytes I/O size (minimum/optimal): 512 bytes / 512 bytes Disk identifier: 0x000c5181 Device Boot Start End Blocks Id System /dev/sda1 * Linux /dev/sda Extended /dev/sda Linux swap / Solaris even brand new hard disks may contain bad blocks mark them format chkdsk create a list of bad low-level formatting and update as needed when a block goes bad, data on that block is lost chkdsk must be manually run to mark it unusable sector sparing/forwarding: maintain a list of spare blocks and replace bad block with one from spare list sector slipping: similar to sector sparing but bad block is replaced by the neighboring block (all blocks are shifted right) 15.6 Swap-Space Management Swap-Space Location Ch 5 and 7: processes in the memory are swapped out to swap memory when available memory is too low Ch 8: virtual memory uses disk space as an extension of main memory, pages are paged out Swap space decreases system performance due to slow disk access time (compared to memory access) Linux suggested swap space: 2 * physical memory With cheaper and larger RAM, swap may not be necessary at all can be out of the normal file system advantage: easy to resize disadvantage: slow due to file-directory traversal e.g.) Windows pagefile.sys more commonly a separate raw partition must be set aside at disk partitioning (OS installation) advantage: faster access disadvantage: internal fragmentation, difficult to resize 6

7 Swap-Space Management: Unix Swap-Space Management: Linux Old Unix: swap out an entire process Solaris 1 (SunOS) text segment pages are thrown page-out, why? stack/heap/uninitialized data of process to swap space Solaris 2 Swap space is allocated only when a page is forced out of physical memory (@ page-out) Kernel uses swap maps to track swap space use similar to Solaris used for stack/heap/uninitialized data of process or shared pages (by multiple processes) Figure Data structures for swapping on Linux systems 11.7 RAID Motivation: disk arrays 11.7 Data Striping data is distributed (striped) across multiple disks Advantages achieve higher data transfer rates on large data access achieve higher I/O rates on small data access better load balancing across disks Disadvantage higher likelihood of disk failures (mean-time-to-failure) Solution provide a form of reliability via redundancy data stripe width: size of data block (>= disk sector size) distributes data over multiple physical disks bit level block level disks, together, appear as a single high-performance disk allow handling of multiple I/O requests in parallel multiple independent I/O requests can be handled in parallel single multiple block requests can be handled by multiple disks in coordination the more disks, the better? 11.7 Data Interleaving Granularity 11.7 RAID fine-grain: small units bit-level: finest all I/O requests access all of the disks in the array advantage: high data transfer rate disadvantage: can handle only one I/O request, all disks have to position (seek+rotate) for every request coarse-grain: large units small I/O requests access only a small # of disks large I/O requests may access all disks in the array advantage: can handle multiple small requests at the same time, higher transfer rate for large requests than single disk Redundant Arrays of Independent/Inexpensive Disks Organization of multiple physical disks into 1 logical unit New disadvantage with redundancy significant performance hit when data changes (disk write) difficult to keep redundant information consistent 7

8 RAID Level 0 striped set (or volume) no redundancy sequential data blocks are written across multiple disks RAID Level 1 mirrored complete data redundancy disk writes can be done in parallel => low write overhead disk reads also in parallel => improve read performance RAID Level 2 bit-level striping with ECC memory-style error correcting-code many modern disks contain own internal ECC RAID Level 3 bit-interleaved parity considers the fact that it is easy to detect which disk failed (which sector of which disk) improved version of RAID 2 (memory-style ECC) 1 parity disk: write, read RAID Level 4 block-interleaved parity similar to bit-interleaved parity (RAID each block-write update block in data disk compute and update parity block in parity disk parity disk is overused and can cause bottleneck RAID Level 5 block-interleaved distributed parity spread data and parity among all disks no one disk is designated as a parity disk all disks contain data and parity for each block, one disk stores parity and others store data parity and data are not stored in the same disk 8

9 RAID Level 6 P + Q redundancy similar to RAID 5 but stores extra redundant information to guard against multiple-disk failures non-binary ECC instead of (binary) parity is used parity can only identify and correct single failure and requires ALL non-failed disks to recover failed disk images from Wikipedia Figure RAID Levels RAID Levels 0+1, 1+0 combinations of RAID levels 0 and 1 RAID 0: provides performance => doubles # of disks RAID 1: provides reliability used to provide both performance and reliability RAID Level 0+1 mirror of stripes resulting stripe is mirrored to another (equivalent stripe) a single disk fails => entire stripe is not accessible wikipedia and textbook RAID Level1+0 stripe of mirrors disks are mirrored in pairs resulting mirrored pairs are striped a single disk fails => its mirror is still available wikipedia and Figure RAID and

10 11.7 Structuring RAID 1. multiple disks connected: OS implements RAID functionality 2. intelligent host controller implements RAID on multiple disks attached to it 3. storage array or RAID array: standalone unit with its own controller, cache, and disks attached to host system via standard controller allows any OS to have RAID-protection 10