Database Reference
In-Depth Information
Server Observed Ingress Bandwidth
900
800
700
600
500
400
300
4 MB
32 MB
64 MB
128 MB
256 MB
200
100
0
0
10
20
30
Number of Processors
40
50
60
70
Server Observed Ingress Bandwidth
1000
900
800
700
600
500
400
300
200
4 MB
32 MB
64 MB
128 MB
256 MB
100
0
0
10
20
30
Number of Processors
40
50
60
70
Figure 5.2
DataTap performance, on the left from the Cray XT3, and on
the right from Infiniband.
a significant reduction in the overhead caused by the data output as the input
set size increases, from about 9% on Lustre to about 3% using DataTap.
*
The structured stream is configured with a simple I/O graph: DataTaps
are placed in each of the GTC processes, feeding out asynchronously to an
I/O node. From the I/O node, each message is forwarded to a graph node
*
We define the I/O overhead as (time with I/O—total time with no I/O)/total time with no I/O.
