-Sandeep Palur and Ajay Anthony (A20302187) (A20306352) 1
Int ntrodu roducti ction n to Clo loudkon udkon Clo loudk udkon on Ar Architectur hitecture Clo loudkon udkon Relo load aded ed Ar Archit hitecture ecture Clo loudk udkon on Rel eloaded aded Imp mprovements rovements Benc nchm hmarkin rking g re resu sults lts Conclus nclusio ion Cont ntrib ribution utions Demo De mo 2
CloudKon is a compact, light-weight, scalable, and distributed task execution framework . Built on following Amazon components: • EC2 • SQS • DynamoDB Major Components in CloudKon: • Client • Server • Global Request Queue (SQS) • Client Response Queue (SQS) 3
4
1. Improved concurrency 2. Bundled Response 3. Efficient Monitoring 5
6
Serve ver Worker rker Thr hread ead (WT) T) 1. Pulls task bundles from global request queue . 2. Creates task thread in optimal concurrency mode. Task sk Thre read ad (TT) T) 1. Deletes the task from the global request queue. 2. Checks for duplication with DynamoDB. 3. Executes task and puts back response to client specific array in Buffer. 7
Buffer ffer (BUF) F): 1. Concurrent hash map Key :Client Response Queue link. Value: ArrayList of task responses. Send end Respo sponse nse Thr hrea ead (SRT) RT): 1. Pulls message bundles from buffer. 2. Sends bundled response to clients. Mon onit itor Thre read ad (MT) T): 1. Attaches object with task thread. 2. Tracks utilization using object’s reference. 8
Client Worker rker Thr hread ead (WT) T): 1. Creates client response queue. 2. Submits task to global request queue. 3. Pulls messages from it’s response queue. 4. Creates task threads using maximum concurrency mode. Task sk Threa read (WT) T): 1. Deletes message from response queue. 2. Adds message in the concurrent ArrayList. 9
1. Improved concurrency • All tasks are processed concurrently. • Reduces Latency. • Increases throughput. 2. Bundled Response: • Reduces network overhead. • Utilizes network bandwidth more effectively i.e. reducing the probablity of network latency. 3. Efficient Monitoring: • Reduces network overhead . • Reduces contention by 1/n, where n = no. of workers 10
Test-be bed: • Ran on Amazon EC2 instances experiments on us.east.1 datacenter of Amazon. • Instance type – m1.large • All instances have Linux OS with JRE 1.7 installed. • Each instance runs both client and server. • 2 client threads and 4 worker threads run on each instance. • Each instance submits 16000 tasks. (8000/thread) • Tasks: sleep 0 , 16, 128
Scripts and programs developed specifically for benchmarking: 1. Shell Scripts (Bash): Throughput, Latency, File transfer from EC2 instances. 2. Parallel-SSH: For parallel execution on EC2. 3. EC2 CLI (Command Line Interface): For instance startup, terminate, Getting IP address, etc. 4. AWS CLI (Command Line Interface): Mainly for Dynamic Provisioning for SQS operations and EC2 dynamic instance startup.
Throughput: • sleep 0 tasks
Throughput Comparison:
Comparison of Sleeps for Throughput: • sleep 0 tasks
Efficiency: • Homogenous workloads. sl sleep ep 128 sleep sl ep 16 100.00% ge) 100% ercentage) ge) ercentage) 80.00% 80% cy (percent cy (percent 60.00% 60% 40% 40.00% sleep 16 sleep 128 ency 20% ency 20.00% icien icien 0% Effici 0.00% Effici 0 100 200 300 0 100 200 300 No. of Instances ces No. of Inst stanc nces es
Consistency: • sleep 16 tasks
Utilization: 8 nodes • sleep 100 tasks 32 tion zation 16 8 iliza Utili 4 8 nodes 2 1 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 Time e (seco econds ) 4 nodes 16 on 8 tion izati 4 iliza Util 4 nodes 2 1 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 Time e (seco econds)
The evaluation of the CloudKon proves that it is highly scalable and achieves a stable performance over different scales. CloudKon achieves up to 87% efficiency. CloudKon was able to outperform other systems like Sparrow and MATRIX on scales of 128 instances or more in terms of throughput. 19
Throughput and efficiency experiments for sleep (0,1,16,128) on the following scales (1,2,4,8,16,32,64,128,256,512,1024). Our code was used for throughput and efficiency benchmarking experiments in CloudKon paper submitted for CCGRID 2014.
21
Questions?? 22
Recommend
More recommend