PROCESSOR ALLOCATION Presented by Wiwek Deshmukh Csc 8320

PROCESSOR ALLOCATION Presented by, Wiwek Deshmukh Csc 8320 Instructor: Dr. Yanqing Zhang

Outline n Allocation Models n Design Issues n Implementation Issues n Example Allocation Algorithms n Processor Allocation in Parallel Databases n References

Processor Allocation n A distributed system consists of multiple processors. Eg: Personal Workstations, A public processor pool, Hybrid Systems. n Some algorithm is required to decide which process should run on which processor.

Allocation Models, Assumptions & Goals n All machines are identical, or atleast codecompatible. n The system is fully interconnected. n When is new work created ? n Processor Allocation Strategies : * Non-Migratory * Migratory - Allow better load balancing but are more complex.

Allocation Models, Assumptions & Goals What are we trying to optimize ? (may differ for different systems) 1. CPU Utilization 2. Mean Response Time 3. Response ratio - Which is better : 1 -sec job that takes 5 secs 1 -min job that takes 70 secs. n

Design Issues Deterministic vs. Heuristic Algorithms. n Centralized vs. Distributed Algorithms. n Optimal vs. Suboptimal Algorithms. n Local vs. Global Algorithms. (Transfer Policy) n Sender-initiated vs. Receiver-initiated Algorithms. (Location Policy) n

Design Issues (Contd …) n n Deterministic algorithms appropriate when everything about process behavior is known in advance. Armed with this information, it is possible to make a perfect assignment. Today’s work is just like Yesterday’s (statistically) When system load is completely unpredictable – we use ad hoc techniques or heuristics.

Design Issues (Contd …) n Centralized vs. Distributed Algorithms. n Scalability and Robustness of Centralized Algorithms. n Centralized Algorithms have been proposed for the lack of suitable decentralized alternatives.

Design Issues (Contd …) n Optimal vs. Suboptimal Algorithms. n Are we trying to find the best solution or an acceptable one ? n Optimal solutions can be obtained in both centralized and decentralized systems. n Optimal solutions are more expensive.

Design Issues (Contd …) n Local vs. Global Algorithms. n What is the transfer policy ? n The choice here is whether or not to base the transfer decision entirely on local information.

Design Issues (Contd …) n Sender-initiated vs. Receiver-initiated Algorithms. n What is the location policy ? n Location policy cannot be local. n Who takes the initiative in locating more CPU cycles ?

Implementation Issues n How do we measure load ? n How do we measure CPU utilization ? n What happens when the kernel executes critical code ? n You may end up underestimating the true CPU usage.

Implementation Issues (Contd …) n n n - n How is overhead dealt with ? Many algorithms ignore the overhead caused by themselves. If system performance improves only by 10% , it may not be worth it. The cost of moving the process may eat up all the gain. The algorithm must consider CPU time, memory usage, and network bandwidth consumed by itself.

Implementation Issues (Contd …) n What is the complexity of the software ? n Eager et al. (1986) studied three algorithms. (Related to location policy) n n Algo-1 : Randomly send to any machine. Algo-2 : Randomly probe any machine. Algo-3 : Probe k machines. n The gain in performance of Algo-3 over Algo-2 is only marginal. n Conclusion of Eager. n

Implementation Issues (Contd …) n Does the stability of the system effect performance ? n Machines run their algorithms asynchronously from one another. n The system is rarely in equilibrium. n What happens when tables are still being updated ?

A Graph-Theoretic Deterministic Algorithm n The goal is to minimize network traffic. n For systems consisting of processes with known CPU and memory requirements. n A known matrix. (Average traffic between each pair of processes) n An assignment to be done for the case: no. of CPUs k < No. of processes for the given goal.

A Graph-Theoretic Deterministic Algorithm (Contd … ) n The system is represented as a weighted graph. n Find a way to partition the graph into k disjoint subgraphs which meet the constraints. n (eg: Total CPU usage & Memory requirements below some limit for each sub-graph. ) n Each sub-graph represents 1 processor.

A Graph-Theoretic Deterministic Algorithm (Contd … ) 3 3 2 USING THE GREEN CUTS : 2 2 1 8 5 6 TOTAL N/W TRAFFIC = 30 4 4 3 (13+17) 5 1 USING THE RED CUTS : 4 2 TOTAL N/W TRAFFIC = 28 (13+15) • Look for Clusters that are tightly coupled.

A Centralized Algorithm (using Heuristic) n A heuristic algorithm does not require any advance information. n Up-Down Algorithm (Mutka and Livny, 1987) n A coordinator maintains a usage table with one entry per workstation. (initially zero) n Allocation decisions are based on the table.

A Centralized Algorithm (using Heuristic) Contd … n This algorithm is designed to distribute computing power more fairly rather than optimizing CPU usage. n When a machine needs a processor, it asks the coordinator. n If no processor is available, the request is denied and a note is made in the table entry.

A Centralized Algorithm (using Heuristic) Contd … n When a machine runs processes on other people’s machines, it accumulates penalty points. (added to table entry) n When it has pending requests, penalty points are subtracted. n Usage table entry can be zero, positive or negative. n Positive score: The workstation is a net user. Negative score: The workstation needs resources. n

A Centralized Algorithm (using Heuristic) Contd … n Heuristic Used: When a processor is free, it is assigned to the workstation having a pending request and a table entry with lowest score. n Thus the intention of the algorithm is to allocate capacity fairly.

A Centralized Algorithm (using Heuristic) Contd … n Disadvantages: 1. Scalability 2. Single point of failure. n Solution: Try to reduce the centralization. or Try semi-centralized techniques.

A Hierarchical Algorithm n These algorithms retain the simplicity of centralized algorithms but scale better. n The processors are organized in a logical hierarchy independent of the physical structure of the network. n Eg: Hierarchy as in an academic institution.

A Hierarchical Algorithm (Contd …) n For each group of k machines, we assign one manager machine. (Dept. head) n If there are too many Dept. heads. , create another level on top. (Deans) n The no. of levels increase logarithmically with the no. of workers.

A Hierarchical Algorithm (Contd …) n Avoid a single (vulnerable) manager at the top. n Truncate the tree at the top to form a committee with ultimate authority. n When a ruling committee member crashes, other members promote someone level down as replacement.

A Hierarchical Algorithm (Contd …) n The system is self-repairing. n Jobs can be created at any level of the hierarchy. n Job with S processes. Allocate R ≥ S processors. R must be large enough but not too large. n

A Hierarchical Algorithm (Contd …) n Disadvantages: 1. Multiple requests are likely to be in various stages of the allocation algorithm. 2. Potential out-of-date estimates. 3. Race conditions & Deadlocks could occur.

A Sender-Initiated Distributed Heuristic Algorithm n Typical algorithms are the ones described by Eager et. al. (1986) n According to their results. , the most cost effective algorithm was Algo-2. n What happens under conditions of heavy load ?

A Receiver-Initiated Distributed Heuristic Algorithm n A complementary algorithm is one initiated by an underloaded receiver. n When a process finishes. , a workstation checks to see if it has enough work. n If not, it begins to randomly probe other machines for work. If work is not found within N probes. , it stops temporarily, and begins again after the next process finishes. n

Comparison of the 2 Distributed Heuristic Algorithms n n SENDER INITIATED: What happens under conditions of heavy load ? Futile attempts made to offload. RECEIVER INITIATED: What happens under conditions of light load ? Futile attempts made to get work.

Comparison of the 2 Distributed Heuristic Algorithms (Contd … ) n Suggested Improvements: n Try to combine both the algorithms. Avoid random polling by keeping a history of past probes. (Try to determine chronic underloaded or overloaded machines). n

A Bidding Algorithm n Proposed by Ferguson et. al. , 1988 n Follows the economics model. n Buyers and Sellers of services and prices set by supply and demand. n The key players are: Processes which buy CPU time & Processors which auction their cycles off to the highest bidder.

A Bidding Algorithm (Contd … ) n Each processor advertises its approximate price by putting it in a publicly readable file. n This price is what the last customer paid. n Different processors have different prices. (Eg: According to speeds. , memory size, Math Co-processors (Floating point hardware etc. ) Services offered may also be published: Eg: Expected Mean response time n

A Bidding Algorithm (Contd … ) When a process wants to start a child process, it checks: n 1. 2. Which processors are offering the services that it needs & which processors can it afford. It then computes the best candidate and makes a bid. n All processors collect all bids and execute the process which made the highest bid. n The published price of the processor is then updated to reflect the going rate.

A Bidding Algorithm (Contd … ) Problems and Potential Questions that arise: n 1. 2. 3. 4. Where do processes get money from ? What is money ? Is disk space also chargeable ? How about laser printer output ? and the list goes on.

An Intelligent Agent for Adaptive Processor Allocation in Parallel Databases B L O C K E D Query Processor 1 db 1 R E A D Y Host Processor 2 db 2 Processor n dbn

An Intelligent Agent for Adaptive Processor Allocation in Parallel Databases PHASE-BASED ALLOCATION: n The query operations are processed in a number of phases. n A set of operations are chosen and allocated to the processors in such a way that the operations complete at approximately the same time. n When all processors finish. , the next phase starts.

An Intelligent Agent for Adaptive Processor Allocation in Parallel Databases NON PHASE-BASED ALLOCATION: n In contrast, this attempts to execute the next awaiting operation as soon as a processor becomes idle. n The awaiting operation(s) is allocated to currently idle processor(s) without waiting for others to complete.

An Intelligent Agent for Adaptive Processor Allocation in Parallel Databases OBSERVATION: n The phase-based algorithm performs better under light load conditions. (Eg: Load < 0. 7) n For a specific load condition, a cross-over point exists in terms of the load. n This can be determined using Supervised Learning. n An intelligent decision can be made as to choose which algorithm under which load condition.

References n Distributed Operating Systems, A. S. Tanenbaum. n Distributed Operating Systems & Algorithms, Randy Chow & Theodore Johnson. n http: //ieeexplore. ieee. org n Processor Allocation Policies for Message-Passing Parallel Computers, Catherine Mccann, Measurement and Modeling of Computer Systems (1994) n An intelligent agent for adaptive processor allocation in parallel databases, Lin, K. H. ; Jiang, Y. ; Leung, C. H. C. ; IEEE International Conference on Intelligent Processing Systems, 1997. ICIPS '97. 1997