I/O, 2002 Journal of Software. Vol.13, No /2002/13(08)

Transcription

1 1-9825/22/13(8) Journal of Software Vol13, No8 I/O, (,184) :, I/O,,, : ; I/O ; ; ; - : TP311 : A,, [1], 1 12 (Terabyte), [2],,,, I/O, [3,4] [5~8],,,, I/O, -, - Exabyte22, : ; : : 973 (G ) : (1969 ),,,,, ; (1947 ),,,,,Web

2 : I/O , I/O ( ):(1) ( zone, zone );(2), ;(3), ;(4), ;(5), ;(6), 1 ;(7), 1 :,,,,,,, Request arrival Input wait queue Drive-1 Drive-2 Drive-n Tape rewind+ drive unload (tape) Robot arm wait queue Robot arm Unload old tape+ load new tape,,, + ( ), +, ( )+ + Fig1 The system model of tape library 1 Drive Drive load (tape)+ seek+transfer data 1, I/O :, I/O, I/O :(1) ;(2) ;(3) 12, 11 11, Exabyte 22 Eliant 82 EXABTYE 8mm, 1 EXABTYE 8mm [5]

3 1614 Journal of Software 22,13(8) Table 1 Measurement of tape library Exabyte 22 1 Exabyte 22 Tape (2) Capacity 7GB(uncompressed ) Tape drive (2) Data transfer rate Mean drive load time Mean drive eject time Rewind startup time Rewind rate Seek startup time Seek rate 92MB/sec (uncompressed) 25s 23s 7s 22MB/s 85s 32MB/s (uncompressed) Tape robot (1) Mean tape unload time 72s Mean tape load time 742s,,,,,,,,, 2, [5~8] - 21 : (1),, ; (2) ( DTL, [3] ) ; (3), (round robin), (MB) ( ) [5], 22 ( ),,,,, ( ), ( ), :,, ;,,

4 : I/O ( ) 2%,,,, :(1), ;, (2) (1), (3),, (1), (2),, 24 -,, ( ),, -, -, ;, - ( ): St Tt Et = Wbw + Wwt (1) S T avg E t t, E t S t t,t t t S avg,t avg,, W bw,w wt, Wwt R w/b ( R w / b = ), ( ), W bw ( ) - -,, - ( ) 3 :(1) COMPAQ -45,Linux ;(2) 11 ;(3), ;(4) 2, 3 :, a/b, a% b%, a>b, b% 3 64M, 7/1, 31 2(a)~(c) ( ) 4 avg

5 1616 Journal of Software 22,13(8),, Static Dynamic Relication-Based Replication-Based-Dynamic ( ), / ( ) ( ( ) ( ) (a) Performance comparison of several scheduling algorithms (light workload) (a) ( ) 45 Static 4 Dynmic 35 Relication-Based 3 Replication-Based-Dynamic ( ), / ( ) ( ( ) ( ) (b) Performance comparison of several scheduling algorithms (moderate workload) (b) ( ) Static Dynmic Relication-Based Replication-Based-Dynamic ( ), / ( ) ( ( ) ( ) (c) Performance comparison of several scheduling algorithms (heavy workload) (c) ( ) Fig2 2 2(a)~(c),,4 (

6 : I/O 1617 <5), (5 <16),, ( 16),, -, -, 3 - ( 2 ), -, - (>15),,, -, Max-Effective-bandwidth-Base d Benefit-Cost-Based scheduling ( ), / Fig3 Benefit-Cost-Based scheduling 3-2(a)~(c), 4,, 4 ( <5),, ;,, ( 25),,, ( 9% ),,, 5( 1) ( 1~1 1/1~1/1),,,,, 4

7 1618 Journal of Software 22,13(8) Improvement ratio (%) Fig Improvement ratio (%) Request skew rate, /, Impact of workload upon dynamic scheduling Fig5 Impact of request skew upon dynamic scheduling (a)~(c),,,,,,,,,, 6( 1) 6,,, 6,,,,, 7 16M 64M Improvement ratio of response time (%) Request skew rate On the head of type In the middle of type In the tail of tape Improvement ratio of response time (%) Request size 16M Request size 64M,, ( ),,, /, 16M, 64M Fig6 Impact of replacement polices of hot Fig7 Impact of request size upon data upon replication-based scheduling replication-based scheduling ,, 2

8 : I/O , >15, Mean response time (s) R w/b, ,, R w/b, 628, , Weight ratio of wait time and effective bandwidth 8( 2, 64M) Fig8 Impact of weight ratio upon benefit- 8, cost-based scheduling 8-4, 9 2 6, 3 1,, experimental 2-simulation 6-experimental 6-simulation Workload(MB/minue): arrival rate*request size ( ), (MB/ ): *, 2-, 2-, 6-, 6- Fig9 Result comparison of simulation and experiment 9 5,,, -,

9 162 Journal of Software 22,13(8) References: [1] Cariño, F, Kaufmann, A, Kostamaa, P Are you ready for Yottabytes? In: Kobler, B, ed Proceedings of the 17th IEEE Symposium on Mass Storage Systems in Cooperation with the 8th NASA GSFC Conference on Mass Storage Systems and Technologies Maryland: IEEE Computer Society Press, 2 476~485 [2] Cariño, F, Burgess, J, O Connell, W, et al Active storage hierarchy, database systems and applications socratic exegesis In: Malcolm, PA, Maria, EO, et al, eds Proceedings of the 25th International Conference on Very Large Data Bases Edinburgh: Morgan Kaufmann Publishers Inc, ~614 [3] Hillyer, BK, Silberschatz, A Random I/O scheduling in online tertiary storage Systems In: Jagadish, HV, Mumick, IS, eds Proceeding of the 1996 ACM SIGMOD International Conference on Management of Data Quebec: ACM Press, ~24 [4] Hillyer, BK, Silberschatz, A Scheduling non-contiguous tape retrievals In: Kobler, B, ed Proceedings of the 15th IEEE Symposium on Mass Storage Systems in Cooperation with the 6th NASA GSFC Conference on Mass Storage Systems and Technologies Maryland: IEEE Computer Society Press, ~124 [5] Hillyer, BK, Rastogi, R, Silberschatz, A Scheduling and data replication to improve tape jukebox performance In: Papazoglou, M, Pu, C, Kitsuregawa, M, eds Proceeding of the 15th International Conference on Data Engineering Sydney: IEEE Computer Society Press, ~541 [6] Nemoto, T, Kitsuegawa, M Scalable tape archiver for satellite image database and its performance analysis with access logs hot declustering and hot replication In: Miller, E, ed Proceedings of the 16th IEEE Symposium on Mass Storage Systems in Cooperation with the 7th NASA GSFC Conference on Mass Storage Systems and Technologies San Diego: IEEE Computer Society Press, ~71 [7] Triantafillou, P, Georgiadis, I Hierarchical scheduling algorithms for near-line tape libraries In: Cammelli, A, Wagner, RR, eds Proceedings of the 1th International Conference and Workshop on Database and Expert Systems Applications Florence: IEEE Computer Society Press, ~54 [8] Prabhakar, S, Dvyakant, A, Amr, El A, et al Scheduling tertiary I/O in database applications In: Roland, W, ed Proceedings of the 8th International Workshop on Database and Expert Systems Applications Toulouse: IEEE Computer Society Press, ~727 Random I/O Scheduling Algorithms in Online Tape Library Systems SHI Jing, ZHOU Li-zhu (Department of Computer Science and Technology, Tsinghua University, Beijing 184, China) {shijing,dcszlz}@mailstsinghuaeducn Abstract: Since the tape libraries have very poor random access performance, it is critical to study random I/O scheduling strategies and algorithms in order to improve the performance of tape library In this paper, the existing scheduling algorithms are summarized first, and then the analytical results of the effectiveness of static scheduling, dynamic scheduling and replication-based scheduling are presented through simulations In particular, a benefit-cost-based scheduling algorithm is given, which aims to improve the effectiveness of existing scheduling algorithms under heavy workloads by tuning the weight ratio of cost and benefit of scheduling policies according to workloads This algorithm is significantly effective under heavy workloads The research of this paper forms the basis of the design of adaptive scheduling algorithms that can be used in massive storage systems Key words: tape library system; random I/O scheduling algorithm; static scheduling; dynamic scheduling; replication-based scheduling; benefit-cost-based scheduling Received February 26, 21; accepted June 13, 21 Supported by the National Grand Fundamental Research 973 Program of China under Grant NoG