BigSimulator (BigNetSim) for Extremely Large Parallel Machines
The BigSim Network Simulator is also known as Bigsimulator and lives in the SVN repository https://charm.cs.uiuc.edu/svn/repos/BigNetSim. The Network simulator is actually more of an Inter-connection network simulator and hence more important in the context of large parallel machines with interconnects. The BigSim simulator along with the network simulator is together also known as BigNetSim.
Both the simulators run on top of the POSE framework, which is a Parallel Discrete Event Simulation framework built on top of Charm++ .
- new types of interconnection topologies and routing algorithms along with different types of switching architecture.
- application performance on different machines. This uses the API provided in Section to run the application on some number of processors on some machine and generate (dump) all events (entry method executions or message send/recv). BigNetSim is used to model the machine that needs to be studied for this application and these logs are then fed into this simulation, and it predicts the performance of this application.
So, the two important uses are studying interconnection networks and performance prediction for applications .
To compile the simulator which is called BigSimulator (or BigNetSim), we need the regular Charm++ build (net-linux-x86_64 in our example). It needs to be complemented with a few more libraries from BigSim and with the Pose discrete-event simulator. These pieces can be built, respectively, with:
./build bgampi net-linux-x86_64 -O2 ./build pose net-linux-x86_64 -O2
Access to the discrete-event simulation is realized via a Charm++ package originally named BigNetSim (now called BigSimulator). Assuming that the 'subversion' (svn) package is available, this package can be obtained from the Web with a subversion checkout such as:
svn co https://charm.cs.uiuc.edu/svn/repos/BigNetSim/
In the subdir 'trunk/' created by the checkout, the file Makefile.common must be edited so that 'CHARMBASE' points to the regular Charm++ installation. Having that done, one chooses a topology in that subdir (e.g. BlueGene for a torus topology) by doing a "cd" into the corresponding directory (e.g. 'cd BlueGene'). Inside that directory, one should simply "make". This will produce the binary "../tmp/bigsimulator". That file, together with file "BlueGene/netconfig.vc", will be used during a simulation. It may be useful to set the variable SEQUENTIAL to 1 in Makefile.common to build a sequential (non-parallel) version of bigsimulator.
- Trace based traffic simulation
- Artificial traffic generation based simulation. The mode of the simulator is governed by the parameter in the netconfig file. When set to 0, trace based simulation is used, when set to 1, traffic generation is used.
./charmrun +p2 ./bigsimulator arg1 arg2
arg1 = 0 => Latency only mode 1 => Detailed contention model arg2 = N => starts execution at the time marked by skip point N (0 is start)
The command line parameters used for this model are different. The format is as follows:
[charmrun +p#] bigsimulator -lat <latency> -bw <bandwidth> [-cpp <cost per packet> -psize <packet size>] [-winsize <window size>] [-skip] [-print_params]
Latency (lat) - type double; in microseconds Bandwidth (bw) - type double; in GB/s Cost per packet (cpp) - type double; in microseconds Packet size (psize) - type int; in bytes Window size (winsize) - type int; in log entries
The implemented equation is:
Latency and bandwidth are required. If cost per packet is given, then packet size must be given, as well. Otherwise, cost per packet defaults to 0.0. Packet size, if given, must be a positive integer.
The -winsize flag allows the user to specify the size of the window (number of log entries) used when reading in the bgTrace log files. This is useful if the log files are large. If -winsize is not specified, the value defaults to 0, which indicates that no windowing will be used (i.e., there will be one window for each time line that is equal to the size of the time line).
As with the second parameter in the examples of part (a) of this section, the -skip flag indicates that the simulation should skip forward to the time stamp set during trace creation (see the BigSim tutorial talk from the 2008 Charm++ workshop). If -skip is not included, then no skipping will occur.
The -print_params flag is provided for debugging convenience. When present, the simple latency model parameters will be displayed during simulation initialization.
./bigsimulator arg1 arg2 arg3 arg4 arg5 arg6
./bigsimulator 1 2 3 100 2031 0.1
arg1 = 0 => Latency only mode 1 => Detailed contention model arg2 = 1 => deterministic traffic 2 => poisson traffic arg3 = 1 => KSHIFT 2 => RING 3 => BITTRANSPOSE 4 => BITREVERSAL 5 => BITCOMPLEMENT 6 => UNIFORM_DISTRIBUTION arg4 = number of packets arg5 = message size arg6 = load factor
- Three dimensional Mesh
- Hybrid of Fattree and Dense Graph
- Hybrid of Fattree and HyperCube
The InitNetwork function must be provided in InitNetwork.C for this new interconnection network. It builds up all the nodes and switches and NICs and channels that form the network. Look at one of the existing interconnection topologies for reference.
This section focuses on the interconnection network simulation. The entities that form an interconnection network are:
A switch decides the routing on a packet. Switches could be
input buffered or output buffered. The former are implemented as individual posers
per port of each switch while the latter are implemented as a poser per switch.
Input Buffered (IB)
switch, a packet in a switch is stored at the input
port until its next route is decided and leaves the switch if it finds
available space on the next switch in the route.
While in an
Output Buffered (OB)
switch, a packet in a switch decides beforehand
on the next route to take and is buffered at the output port until space is
available on the next switch along the route.
Switches are modeled in much detail. Ports, buffers and
virtual channels at ports to avoid head-of-the-line blocking are
modeled. Hardware collectives are implemented on the switch to
enable broadcasts, multicasts and other collective operations
efficiently. These are configurable and can be used if the system
being simulated supports them. We also support configurable
strategies for arbitration, input virtual channel selection and output
virtual channel selection. The configurability of the switch
provides a flexible design, satisfying the requirements of
a large number of networks.
- network card: Network cards packetize and unpacketize messages. A NIC is implemented as two posers. The sending and receiving entities in a NIC are implemented as separate posers. A NIC is attached to each node.
- channel: These are modeled as posers and connect a NIC to a switch or a switch to another switch.
- compute node: Each compute node connects to a network interface card. A compute node simulates execution of entry methods on it. It is also attached to a message traffic generator, which is used when only an interconnection network is being simulated. This traffic generator can generate any message pattern on each of the compute nodes. The traffic generator can send point-to-point messages, reductions, multicasts, broadcasts and other collective traffic. It supports k-shift, ring, bit-transpose, bit-reversal, bit-complement and uniform random traffic. These are based on common communication patterns found in real applications. The frequency of message generation is determined by a uniform or Poisson distribution.
void getNeighbours(int nodeid, int numP);
This is called initially for every switch and this populates the data structure next in a switch which contains the connectivity of that switch. The switch specified by switch has numP ports.
int selectRoute(int current, int dest, int numP, Topology* top, Packet *p, map<int,int> &bufsize, unsigned short *xsubi)
Returns the portid that should be taken on switch current if the destination is dest . The number of ports on a switch is numP . We also pass the pointer to the topology and to the Packet.
int selectRoute(int current, int dest, int numP, Topology* top, Packet *p, map<int,int> &bufsize, map<int,int> &portContention, unsigned short *xsubi)
Returns the portid that should be taken on switch current if the destination is dest . The number of ports on a switch is numP . We also pass the pointer to the topology and to the Packet. Bufsize is the state of the ports in a switch, i.e. how many buffers on each port are full, while portContention is used to give priority to certain ports, when more options are available.
int expectedTime(int src, int dest, POSE_TimeType ovt, POSE_TimeType origOvt, int length, int *numHops)
Returns the expected time for a packet to travel from src to dest , when the number of hops it will need to travel is numHops .
int selectInputVc(map<int,int> &availBuffer, map<int,int> &request, map<int,vector<Header> > &inBuffer, int globalVc, int curSwitch)
Returns the input virtual channel to be used depending on the strategy and the input parameters.