Database Reference
In-Depth Information
design strategy to DataTap described above, and it is an ecient data transfer
substrate that effectively addresses the requirements described above. Unlike
DataTap, which attempts to develop an overall data management framework,
DART is a thin software layer built on RDMA technology to enable fast, low-
overhead, and asynchronous access to data from a running simulation, and
support high-throughput, low-latency data transfers. The design and proto-
type implementation of DART using the Portals RDMA library on the Cray
XT3/XT4 at ORNL are described next. DART has been integrated with the
applications simulating fusion plasma in a Tokamak, described above, and is
another key component of ADIOS.
The primary goal of DART is to eciently manage and transfer large
amounts of data from applications running on the compute nodes of an HPC
system to the service nodes and remote locations, to support remote appli-
cation monitoring, data analysis, coupling, and archiving. To achieve these
goals, DART is designed so that the service nodes asynchronously extract
data from the memory of the compute nodes, and so we ooad expensive
data I/O and streaming operations from the compute nodes to these ser-
vice nodes. DART architecture contains three key components as shown in
Figure 5.3: (1) a thin client layer (DARTClient), which runs on the compute
nodes of an HPC system and is integrated with the application; (2) a stream-
ing server (DARTSServer), which runs independently on the service nodes and
is responsible for data extraction and transport; and (3) a receiver (DARTRe-
ceiver), which runs on remote nodes and receives and processes data streamed
by DARTSServer.
A performance evaluation using the GTC simulation demonstrated that
DART can effectively use RDMA technologies to ooad expensive I/O op-
erations to service nodes with very small overheads on the simulation itself,
allowing a more ecient utilization of the compute elements, and enabling
e
cient online data monitoring and analysis on remote clusters.
DARTClient
DARTSServer
DARTReceiver
P
T
L
P
T
L
P
T
L
P
T
L
T
C
P
SN.1
T
C
P
T
C
P
T
C
P
P
T
L
P
T
L
P
T
L
N.1
N.2
CN.1
CN.2
CN.3
T
C
P
T
C
P
T
C
P
SN.2
P
T
L
P
T
L
P
T
L
T
C
P
T
C
P
CN.4
CN.5
CN.6
N.3
N.4
SN.3
P
T
L
P
T
L
P
T
L
CN.7
CN.8
CN.9
T
C
P
N.5
N.6
SN.4
Asynchronous
Compute node
Service node
Remote cluster
Figure 5.3
Architectural overview of DART.
