Download DIVA: INDEXING HIGH-DIMENSIONAL DATA BY *DIVING* INTO

ALARM: Autonomic Load-Aware Resource
Management for P2P Key-value Stores in Cloud
2011
Dependable, Autonomic and Secure Computing (DASC)
Can Zhang, Haopeng Chen, Shuotao Gao
2015.11.12
분산처리특론
김민수
Index
•
•
•
•
•
•
Introduction
Basic Model
System Design
Experimental Evaluation
Confidence factor
Experimental evaluation
Network & Database Lab
2
Introduction [1/4]
• Relational database problem
 Limitations in scaling up
 Complex data structure
 Poor query performance with SQL
=> Oracle, MySQL, MSSQL, MariaDB etc
• Non-relational databases
 Distributed
 Easily-scalable
 Highly-efficient data store
=> Cassandra, Apache CouchDB, MonggoDB etc
Network & Database Lab
3
Introduction [2/4]
• P2P key-value storage system
 Amazon’s Dynamo
 Google’s Bigtable
 Apache Cassandra
• Consistent Hashing
Network & Database Lab
4
Introduction [3/4]
• P2P key-value storage system
 A lot of existing key-value stores are elastic enough to scale
up or down with no downtime or interruption to application
 Consists of hundreds of machines
 It’s unrealistic for a system administrator to monitor the
system and add/remove a machine manually.
Network & Database Lab
5
Introduction [4/4]
• The contributions of this paper ( ALARM algorithm )
 An autonomic way of resource management for P2P key-value
stores in cloud.
 Resource management algorithm that considers utilization of
multiple resources (CPU, memory, etc.)
 The data stores scale up and down without human
interference.
Network & Database Lab
6
Basic Model [1/4]
Network & Database Lab
7
Basic Model [2/4]
• These nodes participating in the P2P data storage are
defined as virtual nodes.
• Several virtual nodes can be hosted on a physical
node (e.g., a computer or a virtual machine in cloud).
Network & Database Lab
8
Basic Model [3/4]
• The procedure works as follows
 Target : A physical node will become targeted because it’s
either underloaded or overloaded.
Network & Database Lab
9
Basic Model [4/4]
• The procedure works as follows
 Merge/Split : underloaded node -> Merge
overloaded node -> Split
Network & Database Lab
10
System Design [1/2]
Network & Database Lab
11
System Design [2/2]
• The resource management architecture is composed of four
components
 The Storage Control module :
responsible for maintaining a list of available physical nodes in
the data store, and interacting with the Cloud platform to
start/stop a physical node (virtual machine).
 The Status Collector module :
collects the resource utilization on the local machine periodically
(e.g. 20 seconds).
 The Virtual Node :
responsible for data storage.
 The core ALARM module :
The controller in a physical node. Target, Split, Merge.
Network & Database Lab
12
Experimental Evaluation [1/3]
• Five physical servers with hypervisor Xen 3.4.0
Network & Database Lab
13
Experimental Evaluation [2/3]
Network & Database Lab
14
Experimental Evaluation [2/3]
Network & Database Lab
15
Conclusions And Future work
• We present ALARM – an autonomic load – aware
resource management algorithm with respect to the
utilization of multiple resources in the P2P data stores.
• The algorithm detects bottlenecks in the supervised
resources.
• A lot of work has to be done in the future.
Network & Database Lab
16