1. OPERATING SYSTEMS 11 - DISK PIETER HARTEL 1

2. Hardware Intended for humans, local devices and communication (challenge?) Hardware support for interrupts and DMA 2

3. Direct Memory Access (DMA) CPU sends commands to controller indicating: what to do and where in memory to get/put the data Controller runs in parallel with CPU Problem : Contention for system bus cycles Solution : I/O processor or I/O bus 3

4. Design issues Efficiency Buffering (how?) Scheduling Generality Layering Uniform access – Unix! 4

5. #define M 1024*1024 /* One MByte */ int main(int argc, char *argv[]) { int out = open(argv[1], O_RDWR|O_CREAT|O_TRUNC, 0666); int n = atoi(argv[2]); lseek(out, n*M-1, SEEK_SET); write(out, "\0", 1); close(out); return 0; } Sparse files & Network vs disk gcc Sparse.c for d in. /tmp do for n in 1 2 4 8 16 32 64 128 256 512 1024 do time./a.out $d/$n.dat $n time cp $d/$n.dat $d/junk done ls –l /tmp/*dat /tmp.junk 5 Time (s)

6. Mechanisms Disk with rotating platter and moving arm (alternative?) Timing Access time depends on the seek time + rotational delay Transfer time depends on the rotational speed and the channel bandwidth Example: Small blocks? 6 ??? 4ms 4ms 10ms/MB

7. Disk scheduling policies Example: head @ 100, queue contains 55,58,39,18,90,160,150,38,184 Many requests Known address + size Which order? SSTF has an issue… 7

8. Redundant Array of Independent Disks Several physical drives seen as one logical drive Data is striped across the disks (why?) Redundancy (why?) 8

9. Linux design – uniform! Block devices for fast peripherals Character devices for slow peripherals Policies LRU block replacement policy (why?) SCAN scheduler for block R/W Devices are handled as special files: df -l ls –l /dev/cciss/c0d0p2 /dev/tty cat /dev/tty >junk 9

10. #define M 1024*256 #define N 1000 /* number of buffers written */ int main(int argc, char *argv[]) { int out = open(argv[1], O_RDWR|O_CREAT|O_TRUNC, 0666); int i,k; int buf[M]; /* One Mbyte */ for(i=0;i<N;i++) { for(k=0;k<M;k++) { buf[k] = i*N+k; } write(out,buf,sizeof(buf)); } if(argc>=3) { fdatasync(out); posix_fadvise(out,0,0, POSIX_FADV_DONTNEED); } close(out); return 0; } Disk cache man free gcc Fadvise.c 10 CommandCachedTime free -m20705./a.out Foo2.9 free -m21750 cp Foo Foo12.3 free -m22802./a.out Bar x19.8 free -m22778 cp Bar Bar19.3 free -m24829

11. Summary Without I/O no communication with the outside world Caching crucial for performance Disks are slower than the CPU hence more scope for scheduling RAID for more dependable storage 11