1 / 32

Andy Connors

“A cost-based admission control algorithm for digital library multimedia systems storing heterogeneous objects” – I.R. Chen & N. Verma – The Computer Journal – Vol. 46, No. 6, Oct. 2003, pp. 645-659. Andy Connors. Abstract. Multimedia Systems Mixed workloads – Video, Images & Text

Download Presentation

Andy Connors

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. “A cost-based admission control algorithm for digital library multimedia systems storing heterogeneous objects” – I.R. Chen & N. Verma – The Computer Journal – Vol. 46, No. 6, Oct. 2003, pp. 645-659 Andy Connors

  2. Abstract • Multimedia Systems • Mixed workloads – Video, Images & Text • Cost-based admission control algorithm • Based on rewards & penalties • Resource reservation instead of serving requests until all resources exhausted • Reservation based on maximizing total reward • Exploit left over resources • Simulate algorithm and compare to other schemes

  3. Multimedia System

  4. Challenge • Service mixed workloads • Real-time video/audio request – resource demanding and varying data rates • Discrete media – images and text • Need algorithm to “squeeze” in image & text requests without affecting QoS of video requests • However, 70% of data types on Web are image & text

  5. Previous algorithms • Video taking higher priority over image/text data • not justified as 70% of requests are image/text not video • Shenoy & Vin – two-level disk scheduling framework • Level 1: class-independent scheduler – assign bandwidth to application classes – used to dynamically allocate bandwidth to adapt to workload changes – no details on adaption scheme • Level 2: class-specific scheduler – order requests into a common queue for access – minimizes seek time and rotational latency overhead – satisfies QoS requirements of each class – discussed in detail • To & Hamidzadeh – Continious Media-to-Discrete Media redirection ratio • Redirect bandwidth from CM to DM • Allocate more buffer space to CM – reduces admissible CM requests • Optimize disk reads • Use leftover bandwidth for DM requests • How much bandwidth to move from CM to DM requests?

  6. Basic Idea • Dynamically partition resources based on run-time workload changes • Maximize value metric • Ensuring that response time requirements met • Image/text have “own” resources rather than use “leftovers” • Assign value/penalty pair to each request • Value: reward if serviced successfully • Penalty: loss if service rejected due to lack of resources • High value → video higher priority over image/text

  7. Multimedia Server Model • Cycle based disk scheduling: • All requests serviced in TSR – service round duration • Image/text either serviced after video/audio or interleaved – use interleaving to minimize disk seek time and latency • Video/audio requests • As many data blocks as covered by TSR • Double buffered – disk buffer & network buffer • Image/text requests • As many blocks to cover requests object • SCAN algorithm: • Requests ordered and heads traverse in one direction only • Minimizes seek time

  8. Refresher - Scan Algorithm

  9. Resource Partitioning • Text/images serviced in batch • Depart at end of service cycle • Two FIFO queues, one for text, other for images • Statistics of each multimedia object • Distribution of all images and text objects • Histogram of distribution of size needed to satisfy playback • Partition TSR into three parts – video, image and text • Based on cost & workload • Estimate maximum amount of resources allocated to each type • Use left-over time to service more image/text requests

  10. Performance Metric • Maximize reward without compromising QoS (bandwidth & response time) • Reward rate vVNV + vINI + vTNT - qVMV + qIMI + qTMT N{V,I,T} = requests completed per unit time M{V,I,T} = requests rejects per unit time v{V,I,T} = average reward values q{V,I,T} = average penalty values

  11. Algorithm • Use models derived from queing theory • Build lookup table for run-time bandwidth allocation • Estimation of reward rate under given workload condition • Best bandwidth allocation to maximize reward rate • f{V,I,T} = ratio of disk bandwidth for video, image & text requests • fV + fI + fT = 1 (when normalized) • Service times: f{V,I,T}TSR = disk service time • Use statistical admission control to compute number of requests of each type so that probability of disk overload is below a threshold (10-4) • (fV, fI, fT) → (nV, nI, nT) • System behaves like three separate partitions – three queues • For image/text requests • n{I,T} image/text requests per TSR • Total of K{I,T} * n{I,T} image/text requests – K{I,T} = maximum queue size for image/text requests – can use requests in queue to use left-over bandwidth – K{I,T} depends on QoS

  12. Video Request Model • M/M/nV/nV queue • each video stream acts as if served by separate server until departs • V, V = arrival/departure rate of video requests

  13. Video Request Model • Pv(j) = probability that j video out of nV slots occupied • 0 ≤ j ≤ nV • V, V = arrival/departure rate of video requests

  14. Video Request Reward • With probability Pv(j), reward rate = j*vV*V • So total reward gained = jvVV Pv(j) • Rejection rate = V Pv(nV) • Lost reward = qV V Pv(nV) • Reward rate from video = RV RV = (jvVV Pv(j) ) - qV V Pv(nV)

  15. Image & Text Model • For K{I,T}≥1- M/M/1[n {I,T}]/ K{I,T}* n{I,T} queue • Let K{I,T} = 2

  16. Image & Text Model • PI(j) = probability that j video out of nV slots occupied • 0 ≤ j ≤ nI • I, I = arrival/departure rate of video requests • Let KI = 1

  17. Image & Text Model • PI(j) = probability that j video out of nV slots occupied • 0 ≤ j ≤ nI • I, I = arrival/departure rate of video requests • Let KI = 2

  18. Image/Text Request Reward • With probability PI(j) reward rate = • j*vI*I if j < nI • nI*vI*I if j ≥ nI • Rejection rate = I PI(KInI) • Lost reward = qI IPI(KInI) • Reward rate from video = RI RI = ( jvII PI(j) ) + (nIvII PI(j) ) - qI I PI(KInI) j = 1 … nI -1j = nI … KInI

  19. Maximizing Reward • Given V,V,I,I,T,T,vV,qV,vI,qI,vT,qT • Maximize R by searching for optimal (nV, nI, nT) → (n*V, n*I, n*T) • Subject to condition (normalized to text requests) • Here NV, NI, NT are maximum number of requests that can be served of each type (if all bandwidth allocated to each type) • To use total disk bandwidth

  20. Search • Exhaustive • Search all possible solutions • Complexity O(NT2) • Once found all solutions build lookup table • Nearest Neighbor • When NT is too large and exhaustive is computationally too expensive • Complexity O(NT) • Fix one nV, nI, nT then next etc. • Heuristic – largest product of arrival rate and reward selected first

  21. Admission Control Algorithm • Use lookup table to dynamically change to a set of (n*V, n*I, n*T) depending on workload • By monitoring input rates • Use for admission control • Worst case response time for image and text is K{I,T} TSR • Use common schedule queue for disk requests • If total schedule time < TSR use image/text at head of respective queues to use up remaining time by moving to common queue • Probablity that image will be placed on queue f*I/ (f*I+f*T) • And for text f*T/ (f*I+f*T)

  22. Analysis • Numerical analysis of reservation system • Parameters: • Disk Array • 4 disks • Average seek time = 11ms • Rotational latency of 5.5ms • Read/write rate  = 33.3MBps • TSR = 1 • Block size = 4 sectors (512bytes) = 2Kbytes • Images • Evenly distributed across [10kB, 500kB] • Text • Evenly istributed across [1kB, 50kB] • Video • Star Wars – 7200 groups of pictures = 0.5s playback time • 12 frames per group • Calculate • NV = 53, NI = 37, NT = 57 • Simulate • V in range [10,100] arrivals/min, V in range [100,2000], I in range [100,2000]

  23. Other schemes • Compare with other algorithms: • Video First • Highest priority to video requests • Left-overs used for image/text • (nV, nI, nT) = (NV, 0, 0) • Use queue sizes of K{I,T} n*{I,T} • Greedy • Allocates disk in proportion to product of reward and arrival rate • (nV, nI, nT) = ( , , )

  24. Analysis Results

  25. Effect of Arrival Rates • Effect of varying image/text arrival rates as video arrival rate increases • For lower image/text rates • reward rate increases as video rates increase until hit a maximum where we see a decrease • For higher image/text rates • Steadingly decreases due to rejects

  26. Effect Of Video Departure Rate • Using varying video departure rates shows effect on increasing video arrival rate • At higher departure rates • See an increase in reward rate as arrival rate increases until a threshold where server is heavily loaded and rejects requests • At lower • Video requests stay in system for longer time and so system admits fewer requests

  27. Effect Of Video Reward Value • Using varying video reward values shows effect on increasing video arrival rate • At higher reward rates • Systems admits more requests – threshold shifts higher

  28. Results – Reward Rate • Under light loads • Close to predicted lower-bound reward rates • At higher loads • Higher than calculated – due to effect of using left-over bandwidth which is more pronounced at higher loads • In limit • Returns back to theoretical as text/image queues are full and consume all server resources • Same as video-first at lower loads • as system can accommodate most users at these loads • At higher loads • Out performs both video-first and greedy algorithms

  29. Results – Response Time • Under light loads • Close to other algorithms • At higher loads • As explicitly allocate time for image/text request see better response times than video-first – difference between 1s and 5s • Greedy favors video/text and so has better response times – but compares favorably

  30. Results – Utilization • Does not show greedy algorithm as shows same trends as reservation algorithm • For video-first • Higher utilization for video requests – lower for image/text • For reservation • Better utilization for image/text • Lower for video

  31. Results – Rejection Rates • At higher loads • Rejects fewer image/text requests than video-first or greedy • Achieved by rejecting more video requests • Video-first rejects 0 video requests but a high number of image/text

  32. Conclusions • Significant improvement in reward rate compared to video-first and greedy algorithms • Without sacrificing performance metrics such as response time & system utilization

More Related