Storing and Serving Multimedia

What is a Media Server? A scalable storage manager Allocates multimedia data optimally among disk resources Performs memory and disk-based I/O optimization Supports real-time and non-real-time clients Supports presentation of continuous-media data Supports mixed workloads – schedules the retrieval of blocks Performs admission control

Task 1 How should multimedia data be placed on a disk?

Constrained Allocation Intuition: presentations are sequential, so place sequential segments of a MM object as closely as possible MM object X[i] = {X[i,1], X[i,2] …X[i,n]} | Track(X[i,j+1]) - Track(X[i,j]) |  d Problem?

Scheme Bidirectional Disk has R equal regions Max length of seek= 2/R X disk radius What combination of R and segment size S will maximize throughput N without violating continuity and memory availability? Time between two I/O requests = T, T  S/DR For N requests, total time T = T regionTR + T regionSeek Time to transfer N segments T regionTR = N  (S/TR) Worst seek = d = CYL/R tracks   (CYL/R) with overhead T regionSeek = N   (CYL/R) N  (S/TR +  (CYL/R) )  S/DR What is wrong with this model? data display rate data transfer rate  (d)=a1+b1  d, often piecewise continuous

Scheme Bidirectional T is never constant The i-th I/O request has to wait till the (i-1)th request has been serviced + seek time + transfer time Time between two consecutive requests: T + i   (CYL/R) Solution Increase the segment by DR  i   (CYL/R) for the i-th request For N requests What is the worst initial latency under this model?

Other Allocation Schemes Unidirectional allocation Initial Latency Sequential allocation Idea: pick up whatever segments in on-disk sequential order Expected seek distance: CYL/(R X N) Worst case: 2(N-1) accesses between segments X[i,j] and X[i,j+1] Solution Put X[i,1] into a buffer (size S) and start playback only after the region k containing X[i,1] is fully read If X[i,2] occurs at the start of region k+1, put in a cushion buffer, else if occurs at the end of region k+1play back from disk. Memory constraint: 2  N  S  Mem Concurrent presentations

Other Allocation Schemes Arrival Time Problem: X[i] comes just when the disk passes over X[i,1] Wait period: R region scans + time to read region k Group Sweeping Scheme (GSS) Divide N requests into G groups Service N/G requests as a chunk When is this beneficial? How does initial latency of the scheme compare with others? Effects of Replication Memory cost? Disk cost? Latency?

Service Models Random Access Maximize the number of clients that can be served concurrently at any time with a low response time Minimize latency Enhanced Pay-per-view (EPPV) Increase the number of clients that can be serviced concurrently beyond the available disk and memory bandwidth, while guaranteeing a constraint on the response time