Download presentation
Published byOsborne McLaughlin Modified over 9 years ago
1
The IEEE International Conference on Cluster Computing 2010
CDRM: A Cost-effective Dynamic Replication Management Scheme for Cloud Storage Cluster Qingsong Wei Data Storage Institute, A-STAR, Singapore Bharadwaj Veeravalli, Bozhao Gong National University of Singapore Lingfang Zeng, Dan Feng Huazhong University of Science & Technology, China
2
Cost-effective Dynamic Replication Management (CDRM)
Agenda Outline Introduction Problem Statement Cost-effective Dynamic Replication Management (CDRM) Evaluation Conclusion
3
1, Introduction Outline HDFS Architecture Clients Name Node Network
Meta Data Name Node Network Control Node Disk Disks Node Disk Disks Node Disk Disks Node Disk Disks Node Disk Disks Node Disk Disks Data Blocks Data Nodes
4
1. Introduction In the HDFS, files are striped into date blocks across multiple data nodes to enable parallel access. B1 B2 … Bm Data Striping However, Block may be unaccessible due to date node unavailable. If one of the blocks is unavailable, so as the whole file. Failure is normal instead of exception in large scale storage cloud system. Fault tolerance is required in such a system. … Node1 Node2 Noden Page 4 of 19
5
1. Introduction Data nodes Clients 1 1 2 5 2 2 4 3 4 5 4 3 5
Replication is used in HDFS. When one data node fails, the data is still accessible from the replicas and storage service need not be interrupted. Besides fault tolerance, replicas among data nodes can be used to balance workload. Page 5 of 19 5
6
2. Problem Statement Current replication managements
Treat all data as same: same replica number for all data Treat all storage nodes as same Fixed and Static 1 2 3 4 5 5 1 2 3 4 4 5 1 2 3 High cost & Poor load balance Page 6 of 19 6
7
2. Problem Statement Replica number is critical to management cost. More replica, more cost. 1 2 1 5 2 2 4 The block 5 is modified 3 5 4 4 3 5 Update to maintain consistency Because large number of blocks are stored in system, even a small increase of replica number can result in a significant increase of management cost in the overall system. Then, how many minimal replica should be kept in the system to satisfy availability requirement? Page 7 of 19 7
8
2. Problem Statement Replica placement influences intra-request parallelism. Client File (B1, B2, B3) B3 B2 B1 Requests Blocked If B1 has a replica in Datanode 2 which has free sessions, the access can be immediately serviced. B3 B2 B1 B1 Data Node1 Data Node2 Data Node3 Sessionmax=3 Sessionfree=1 Sessionmax=3 Sessionfree=2 Sessionmax=2 Sessionfree=0 Page 8 of 19
9
2. Problem Statement Replica placement also influences inter-request parallelism. Client1 Client2 B3 B2 B1 B1 Requests How to place these replicas among Data nodes clusters in a balance way to improve access parallelism? B1 B2 B1 B3 Data Node1 Data Node2 Data Node3 Sessionmax=3 Sessionfree=0 Sessionmax=3 Sessionfree=1 Sessionmax=2 Sessionfree=0 Page 9 of 19
10
3. Cost-effective Dynamic Replication Management
System Model pj : popularity sj : size rj : replica number tj : access latency requirement (p1, s1, r1, t1) (pj, sj, rj, tj) (pM, sM, rM, tM) …… …… B1 Bj BM Total arrival rate: λ λi : req. arr. rate τi : average ser. time fi : failure rate ci : max sessions Node1 … Nodei … NodeN Two key issues to be addressed 1, Minimal replica number 2, How to place these replicas in balanced way (λ1, τ1, f1, c1) (λi, τi, fi, ci) (λN, τN, fN, cN) Data has different attributes Data nodes are different Page 10 of 19
11
3. Cost-effective Dynamic Replication Management
Availability Suppose file F is striped into m blocks {b1 , b2 ,…, bm}. To retrieve whole file F, we must get all the m blocks. Availability is modeled as function of replica number. Suppose the expected availability for file F is Aexpect, which defined by users. To satisfy the availability requirement for a given file, we get Minimum replicas can be calculated from above Eq. for a given expected availability. Page 11 of 19
12
3. Cost-effective Dynamic Replication Management
Blocking Probability Blocking probability is used as criterion to place replicas among data nodes to improve load balance . An data node Si is modeled as M/G/ci system with arrival rate λi and service time τi, and accordingly, the blocking probability of data node Si can be given to be Replica placement policy: replica will be placed into data node with lowest blocking probability to dynamically maintain overall load balancing. After determining how many replicas the system should maintain at least to satisfy availability requirement, we shall explain how to place these replicas efficiently to maximize performance and load balancing. Blocking probability describe the probability of data node to reject a request. Page 12 of 19
13
Framework of cost-effective dynamic replication management in HDFS
Request to create a file with <Availability, Block Number> 1 Client Name Node Bm … B2 B1 Return replication policy <Bi, Replication factor, DataNode list> 3 4 Flush and replicate blocks to selected Datanodes Calculate the replication factor and Search the Datanode B+Tree to obtain Datanode list. 2 Data Nodes Besides default replication management, HDFS provides flexible API for us to extend and implement efficient replication management. Replication Pipelining Framework of cost-effective dynamic replication management in HDFS Page 13 of 19
14
4. Evaluation Setup Our test platform is built on a cluster with one name node and twenty data nodes of commodity computer The operating system is Red Hat AS4.4 with kernel Hadoop version is and java version is AUSPEX file system trace is used A synthesizer is developed to create workloads with different characteristics, such as data sets of different sizes, varying data rates, and different popularities. These characteristics reflect the differences among various workloads to the cloud storage cluster. Page 14 of 19
15
4. Evaluation Cost effective Availability
Initially, one replica per object. CDRM only maintain minimal replicas to satisfy availability. Higher failure rate, more replica required. Dynamic replication with Data node failure rate of 0.1 and 0.2 , Aexpect=0.8 Page 15 of 19
16
Effect of popularity and access arrival rate, 20 data nodes
4. Evaluation Performance CDRM vs. HDFS default Replication Management (HDRM) under different popularity and workload intensity. Performance of CDRM is much better than that of HDRM when popularity is small. CDRM outperform HDRM under heavy workload. popularity is a ratio between the size of the most popular data and the size of the total data set. A large popularity indicates the data are more uniformly accessed. A small popularity means accesses concentrate in a small amount of data. Effect of popularity and access arrival rate, 20 data nodes Page 16 of 19
17
System utilization among data nodes, popularity=10%, λ=0.6
4. Evaluation Load Balance The figure shows the difference of system utilization of each data node comparing to the average system utilization of the cluster. CDRM can dynamically distribute workload among whole cluster. System utilization among data nodes, popularity=10%, λ=0.6 Page 17 of 19
18
Current replication management policies
5. Conclusion Current replication management policies CDRM Data is same Data is different Storage node are same Storage nodes are different Same replica number for all data Different replica number for different data Static placement Dynamic placement High Cost Cost effective Poor load balance Good balance Low performance High performance Page 18 of 19
19
Thanks & Question For more questions, please contact Dr. Qingsong Wei by Page 19 of 19
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.