On The Scalability of Storage Sub-System Back-end Network

Transcription

1 n he calability of torage ub-ystem ack-end etwork an Li, oland bbett, igel opham and im Courtney chool of nformatics niversity of dinburgh Xyratex, K an Li n he calability of torage ub-ystem ack-end etwork.li-24@sms.ed.ac.uk

2 Project Motivation 1 Project Motivation heoretically the more disks used in a disk array, the higher the degree of parallesim, so leading to larger the performance potential performance benefits. owever, in a real system there is a limitation on the scale of A systems due to the limitation of interconnection network. he more disks are added to the system, the higher the contention for the shared media. When the and cache size in a A system reaches a certain threshold, there will be no further gain in performance by adding more disk or cache due to the saturation of the back-end network. an Li n he calability of torage ub-ystem ack-end etwork.li-24@sms.ed.ac.uk

3 Project oal 2 Project oal nvestigate the capacity of interconnection networks in terms of the numbers of disks that can be included in one chain. n particular, ibre Channel (C) is chosen as the interconnection network. ive a certain and cache size, how much bandwidth is necessary to support them to get the maximum performance? Likewise, how many disks (and cache) can be connected to a 2C (4C) port? an Li n he calability of torage ub-ystem ack-end etwork.li-24@sms.ed.ac.uk

4 Analytical Models 3 Assumptions: he request size of random access workload is equal to the size of stripe unit. he access address of each request is aligned to the stripe unit boundary. he capacity of the A system keeps fixed for all the study. he queue length of disk is 1, ie., no disk command waits in the disk for service. he C is chosen as the research subject. isk command transmission time x = + overhead.. isk command execution time d = port + size of stripe unit; network bandwidth; media + seek. an Li n he calability of torage ub-ystem ack-end etwork.li-24@sms.ed.ac.uk

5 Analytical Models 4 Large equential Access v x d x v = x+ d ime or sequential access workload, a large volume command is divided into disk commands, so that the response time for that command v = x + d. hroughput is the major performance metric: hroughput = K v = K x + d. an Li n he calability of torage ub-ystem ack-end etwork.li-24@sms.ed.ac.uk

6 Analytical Models 5 mall andom Access x v d x v d ( 1)x ( 1)x ( 1)x < d (= 3) ime ( 1)x> d v = x+ d v = x (= 5) ime x + d ( 1) x < d v = x ( 1) x >= d or random access workload, P is the major performance metric, P = x + d d x 1 x ( 1) x < d ( 1) x >= d an Li n he calability of torage ub-ystem ack-end etwork.li-24@sms.ed.ac.uk

7 Analytical Models 6 Analytical esults (no cache) 16 x eq ead, K=128Kytes 8 throughput (ytes/s) 1 8 P 6 AM, =16Kytes equential Access andom Access an Li n he calability of torage ub-ystem ack-end etwork.li-24@sms.ed.ac.uk

8 Analytical Models 7 eneral Model = (,C,, L,P) = um d + overhead, the bandwidth required to achieve the maximum performance with disks and C cache in system. : C: cache size : size of stripe unit L: workload characteristic P: cache destage threshold, P= for our study um d : number of disk commands send to disk per second an Li n he calability of torage ub-ystem ack-end etwork.li-24@sms.ed.ac.uk

9 imulation esults 8 ystems without Cache ( = 16K) 8 1 Max PC A5 A6 utilization isk ti A5 isk ti A6 et ti A5 et ti A Port andwidth = bps Max PC A5 A6 utilization isk ti A5 isk ti A6 et ti A5 et ti A Port andwidth = 4.25 bps an Li n he calability of torage ub-ystem ack-end etwork.li-24@sms.ed.ac.uk

10 imulation esults 9 ystems with Cache ( = 32K) Max PC (a) A5 A6 bandwidth (bps) (b) A5 A (c) A5 A6.75 (d) A5 A6 disk utilization.9.85 read miss rate an Li n he calability of torage ub-ystem ack-end etwork.li-24@sms.ed.ac.uk

11 ummary 1 ummary When this is no cache, a 2 C port is able to support up to 46 disks for A5 and 53 disks for A6 (size of stripe unit = 16kytes). With enough cache, a 2 C port is able to support up to 18 disks under LP like workload. (size of stripe unit = 32Kytes). When there is enough cache, the bandwidth required to support a certain is fixed. t is irrelevant with protection level and cache size. an Li n he calability of torage ub-ystem ack-end etwork.li-24@sms.ed.ac.uk

12 uture Work 11 uture Work tudy the scalability of back-end network when the size of stripe unit is 16k and the system performance. tudy the network bandwidth requirement when there is cache coherency between two controllers. an Li n he calability of torage ub-ystem ack-end etwork.li-24@sms.ed.ac.uk