DDFD_Performance_Poster_01

Download Report

Transcript DDFD_Performance_Poster_01

Annual Conference of ITA
ACITA 2009
Experimental Evaluation of the Performance and Scalability of
a Dynamic Distributed Federated Database
Graham Bent (IBM UK) , Patrick Dantressangle (IBM UK), Abbe Mowshowitz (CUNY), Paul Stone (IBM UK), David Vyvyan (IBM UK)
Experimental evaluation confirms that query ‘execute time’ is a function of the
number of nodes (N) and the network latency (TL) between nodes and scales
as O(TL logN). Data ‘fetch time’ is linear with number of records returned.
1
2 A 1025 Node Gaian Database
Experimental Testbed
What was investigated:
Physical architecture:
•
22 IBM Blade Servers distributed over two Blade Centres,
plus 3 Logical Partitions of an IBM System P Server;
• Gigabit Ethernet interconnection;
• Constructed databases of various sizes up to 1250 nodes.
3
•
The time to propagate a query to all of the nodes in the database,
as a function of the number of database nodes (N);
• The time to fetch data from across the nodes of the database to a single node
as a function of the volume of data;
• The time to fetch data from multiple nodes concurrently queries
as a function of the number of nodes concurrently querying.
Experimental Results
Query time to fetch 1 million rows
Individual Query Time Scalability
6000
592.9
Total Query Time 1025 nodes
539.0
Total Query Time 1 node
5000
Total Qury Time 1 node indexed
485.1
431.2
Time (milliseconds)
4000
377.3
Query Time (ms)
Linear (Total Query Time 1025
nodes)
Linear (Total Query Time 1 node)
323.4
y = 4.217x + 349.251
3000
269.5
2000
215.6
Average Query Time
161.7
y = 1.7383x + 678.141
Predicted Max (Diameter+1)
Predicted Min (Radius+1)
107.8
1000
Queried node eccentricity+1
53.9
0
0.0
0
200
400
600
Number of Nodes
Query Execution Time:
• Average Query Execute Time
with TL=50ms;
process
Copy delay Tp= 3.9ms
for database size up to 1025 nodes.
800
1000
1200
0
200000
400000
600000
800000
1000000
1200000
Total Rows fetched
Data Fetch Time:
Query Fetch Time and Execute time
as a function of the number of rows fetched per node from
a 1025 node database.