Open Versus Closed: A Cautionary Tale

Open Versus Closed:A Cautionary Tale 報告者: R02944041楊名揚 R02944032鄭琮錡 B98902023陳柏廷

Abstract-Three purposes • how scheduling policies are impacted by closed and open models, and explain the differences in user level performance. • the use of partly open system models, whose behavior we show to lie between that of closed and open models. • determining which system model is most appropriate for a given workload.

Introduction-Three systems

Introduction • The builders often do not seem to view this as an important factor worth mentioning in the document • We show that closed and open system models yield significantly different results, even when both models are run with the same load and service demands.

Differences between open and closed • For a fixed load, the mean response time for an open system model can exceed that for a closed system model by an order of magnitude or more. • The impact of scheduling on improving system performance.

Closed systems • number of users is called the multiprogramming level (MPL) and denoted by N • 2 steps：(a) submit a job (b)receive the response and then “think” for some amount of time. • a new request is only triggered by the completion of a previous request. • Nthink + Nsystem = N.

Open systems • A request completion does not trigger a new request: a new request is only triggered by a new user arrival.

Partly-open system • probability p the user stays and makes a follow up request (possibly after some think time), and with probability 1− p the user simply leaves the system

Scheduling policies • FCFS (First-Come-First-Served) • PS (Processor-Sharing) • PESJF (Preemptive-Expected-Shortest-Job-First) • SRPT (Shortest-Remaining-Processing-Time-First) • PELJF (Preemptive-Expected-Longest-Job-First)

Real-world case studies • Static web content ： i.e. requests of the form “Get me a file,” in a LAN environment. • E-commerce site ：the database back-end server of an e-commerce site, e.g. an online bookstore. • Auctioning web site ： This case study uses simulation based on a trace from one of the top-ten U.S. online auction sites.

Real-world case studies

Study ofWAN effects

Open versus closed systems • FCFS • The impact of scheduling

FCFS-Principle(i) • Principle (i): For a given load, mean response times are significantly lower in closed systems than in open systems.

Principle (i) • The open system will always serve as an upper bound for the response time of the closed system • MPL (Multi-Programing Level)

FCFS-Principle (ii) • Principle (ii): As the MPL grows, closed systems become open, but convergence is slow for practical purposes.

FCFS-Principle (iii) • Principle (iii): While variability has a large effect in open systems, the effect is much smaller in closed systems.

Principle (iii) • For an open system, we see that directly affects mean response time. • short jobs were stuck behind long jobs, increasing mean response time. • For the closed system with MPL 10, has comparatively little effect on mean response time.

The impact of scheduling-Principle(iv) • Principle (iv): While open systems benefit significantly from scheduling with respect to response time, closed systems improve much less.

Principle(iv) FCFS :First-Come-First-Served PS: Processor-Sharing (similar to Round Robin) PESJF: Preemptive-Expected-Shortest-Job-First SRPT: Shortest-Remaining-Processing-Time-First PELJF: Preemptive-Expected-Longest-Job-First

Principle(iv) • For an open system, scheduling can prevent small jobs from queueing behind large jobs. • For the close system, scheduling is not well understood.

The impact of scheduling-Principle(v) • Principle (v): Scheduling only significantly improves responsetime in closed systems under very specific parametersettings: moderate load (think times) and highMPL.

The impact of scheduling-Principle(vi) • Principle (vi): Scheduling can limit the effect of variability in both open and closed systems. • Better scheduling (PS and PESJF) helps combat the effect of increasing variability

Partly-open systems • Session • Principle (vii) • Principle (vii)

Session • Aset of requests for each customer

Partly-open system-principle (vii) • Principle (vii): A partly-open system behaves similarly to an open system when the expected number of requests per session is small and similarly to a closed system when the expected number of requests per session is large. • 當requests per session越大,feedback的數量越多,越像closedsystem

Partly-open system-principle (viii) • Principle (viii): In a partly-open system, think time has little effect on mean response time.

Choosing a system model(I) • Modeling?

Choosing a system model(II) 1. Collect traces from the system. 2. Construct a partly-open model for the system, since the partly-open model is the most general and accurate. In particular, obtain the relevant parameters for the partly-open model. 3. For the partly-open model, decide whether an open or a closed model is appropriate, or if the partly-open model is necessary.

Choosing a system model(III) Principle (vii) + Principle (viii) ---------------------------------- number of request / (session)[8,26] Def: user ,session, timeout . 1. Defacto standard value ->1800s X 2.

Choosing a system model(IV)

PriorWork • Service variability dominant: open>close • FCFS-----MPL↑ open ←→closed

Conclusion(I) • 1.response time close < open (MPL↑) close = open • 2.(a)gap of open&close in moderate load (b)variability↑ ->convergence rate↓ (c)scheduling benefits: open>close • (Scheduling effective in open) • (close for moderate load and high variability)

Conclusion(II) • 3.(a)Lot simultaneously users.(<1000) ->open (b)Lot request per sessions.(<10) ->close • 4.Both affected by variability. ->high variability for larger cutoffs • 5.Think time only effect the load.

Thanks for listening

Open Versus Closed: A Cautionary Tale