NAME¶
NetPIPE -
Network Protocol Independent
Performance Evaluator
SYNOPSIS¶
NPtcp [
-h receiver_hostname]
[
-b TCP_buffer_sizes] [options]
mpirun [
-machinefile hostlist] -np 2
NPmpi [-a] [-S]
[-z] [options]
mpirun [
-machinefile hostlist] -np 2
NPmpi2 [-f]
[-g] [options]
NPpvm [options]
See the TESTING sections below for a more complete description of how to run
NetPIPE in each environment. The OPTIONS section describes the general options
available for all modules. See the README file from the tar-ball at
http://www.scl.ameslab.gov/Projects/NetPIPE/ for documentation on the
InfiniBand, GM, SHMEM, LAPI, and memcpy modules.
DESCRIPTION¶
NetPIPE uses a simple series of ping-pong tests over a range of message
sizes to provide a complete measure of the performance of a network. It
bounces messages of increasing size between two processes, whether across a
network or within an SMP system. Message sizes are chosen at regular
intervals, and with slight perturbations, to provide a complete evaluation of
the communication system. Each data point involves many ping-pong tests to
provide an accurate timing. Latencies are calculated by dividing the round
trip time in half for small messages ( less than 64 Bytes ).
The communication time for small messages is dominated by the overhead in the
communication layers, meaning that the transmission is latency bound. For
larger messages, the communication rate becomes bandwidth limited by some
component in the communication subsystem (PCI bus, network card link, network
switch).
These measurements can be done at the message-passing layer (MPI, MPI-2, and
PVM) or at the native communications layers that that run upon (TCP/IP, GM for
Myrinet cards, InfiniBand, SHMEM for the Cray T3E systems, and LAPI for IBM SP
systems). Recent work is being aimed at measuring some internal system
properties such as the memcpy module that measures the internal memory copy
rates, or a disk module under development that measures the performance to
various I/O devices.
Some uses for NetPIPE include:
Comparing the latency and maximum throughput of various network cards.
Comparing the performance between different types of networks.
Looking for inefficiencies in the message-passing layer by comparing it to the
native communication layer.
Optimizing the message-passing layer and tune OS and driver parameters for
optimal performance of the communication subsystem.
NetPIPE is provided with many modules allowing it to interface with a
wide variety of communication layers. It is fairly easy to write new
interfaces for other reliable protocols by using the existing modules as
examples.
TESTING TCP¶
NPtcp can now be launched in two ways, by manually starting NPtcp on both
systems or by using a nplaunch script. To manually start NPtcp, the NetPIPE
receiver must be started first on the remote system using the command:
NPtcp [options]
then the primary transmitter is started on the local system with the command
NPtcp -h
receiver_hostname [options]
Any options used must be the same on both sides.
The nplaunch script uses ssh to launch the remote receiver before starting the
local transmitter. To use rsh, simply change the nplaunch script.
nplaunch NPtcp -h
receiver_hostname [options]
The
-b TCP_buffer_sizes option sets the TCP socket buffer
size, which can greatly influence the maximum throughput on some systems. A
throughput graph that flattens out suddenly may be a sign of the performance
being limited by the socket buffer sizes.
TESTING MPI and MPI-2¶
Use of the MPI interface for NetPIPE depends on the MPI implementation being
used. All will require the number of processes to be specified, usually with a
-np 2 argument. Clusters environments may require a list of the hosts
being used when each job is run. Put the list of hosts in hostlist then, for
OpenMPI, run NetPIPE using:
mpirun --hostfile
hostlist -np 2 NPmpi [NetPIPE options]
For MPICH2 use instead:
mpirun -machinefile
hostlist -np 2 NPmpi [NetPIPE options]
To test the 1-sided communications of the MPI-2 standard, compile using:
make mpi2
Running as described above and MPI will use 1-sided MPI_Put() calls in both
directions, with each receiver blocking until the last byte has been
overwritten before bouncing the message back. Use the
-f option to
force usage of a fence to block rather than an overwrite of the last byte. The
-g option will use MP_Get() functions to transfer the data rather than
MP_Put().
TESTING PVM¶
Start the pvm system using:
pvm
and adding a second machine with the PVM command
add
receiver_hostname
Exit the PVM command line interface using quit, then run the PVM NetPIPE
receiver on one system with the command:
NPpvm [options]
and run the TCP NetPIPE transmitter on the other system with the command:
NPpvm -h
receiver hostname [options]
Any options used must be the same on both sides. The nplaunch script may also be
used with NPpvm as described above for NPtcp.
TESTING METHODOLOGY¶
NetPIPE tests network performance by sending a number of messages at each
block size, starting from the lower bound on the message sizes.
The message size is incremented until the upper bound on the message size is
reached or the time to transmit a block exceeds one second, which ever occurs
first. Message sizes are chosen at regular intervals, and for slight
perturbations from them to provide a more complete evaluation of the
communication subsystem.
The
NetPIPE output file may be graphed using a program such as
gnuplot(1). The output file contains three columns: the number of bytes
in the block, the transfer rate in bits per second, and the time to transfer
the block (half the round-trip time). The first two columns are normally used
to graph the throughput vs block size, while the third column provides the
latency. For example, the
throughput versus block size graph can be
created by graphing bytes versus bits per second. Sample
gnuplot(1)
commands for such a graph would be
set logscale x
plot "np.out"
OPTIONS¶
- -a
- asynchronous mode: prepost receives (MPI, IB modules)
- -b TCP_buffer_sizes
- Set the send and receive TCP buffer sizes (TCP module only).
- -B
- Burst mode where all receives are preposted at once (MPI, IB modules).
- -f
- Use a fence to block for completion (MPI2 module only).
- -g
- Use MPI_Get() instead of MPI_Put() (MPI2 module only).
- -h hostname
- Specify the name of the receiver host to connect to (TCP, PVM, IB, GM).
- -I
- Invalidate cache to measure performance without cache effects (mostly
affects IB and memcpy modules).
- -i
- Do an integrity check instead of a performance evaluation.
- -l starting_msg_size
- Specify the lower bound for the size of messages to be tested.
- -n nrepeats
- Set the number of repeats for each test to a constant. Otherwise, the
number of repeats is chosen to provide an accurate timing for each test.
Be very careful if specifying a low number so that the time for the
ping-pong test exceeds the timer accuracy.
- -O source_offset,dest_offset
- Specify the source and destination offsets of the buffers from perfect
page alignment.
- -o output_filename
- Specify the output filename (default is np.out).
- -p perturbation_size
- NetPIPE chooses the message sizes at regular intervals, increasing them
exponentially from the lower boundary to the upper boundary. At each
point, it also tests perturbations of 3 bytes above and 3 bytes below each
test point to find idiosyncrasies in the system. This perturbation value
can be changed using the -p option, or turned off using -p
0 .
- -r
- This option resets the TCP sockets after every test (TCP module only). It
is necessary for some streaming tests to get good measurements since the
socket window size may otherwise collapse.
- -s
- Set streaming mode where data is only transmitted in one direction.
- -S
- Use synchronous sends (MPI module only).
- -u upper_bound
- Specify the upper boundary to the size of message being tested. By
default, NetPIPE will stop when the time to transmit a block exceeds one
second.
- -z
- Receive messages using MPI_ANY_SOURCE (MPI module only)
- -2
- Set bi-directional mode where both sides send and receive at the same time
(supported by most modules). You may need to use -a to choose
asynchronous communications for MPI to avoid freeze-ups. For TCP, the
maximum test size will be limited by the TCP buffer sizes.
FILES¶
- np.out
- Default output file for NetPIPE. Overridden by the -o
option.
AUTHOR¶
The original NetPIPE core plus TCP and MPI modules were written by Quinn Snell,
Armin Mikler, Guy Helmer, and John Gustafson. NetPIPE is currently being
developed and maintained by Dave Turner with contributions from many students
(Bogdan Vasiliu, Adam Oline, Xuehua Chen, and Brian Smith).
Send comments/bug-reports to:
<netpipe@scl.ameslab.gov>.
Additional information about
NetPIPE can be found on the World Wide Web
at
http://www.scl.ameslab.gov/Projects/NetPIPE/
BUGS¶
As of version 3.6.1, there is a bug that causes NetPIPE to segfault on RedHat
Enterprise systems. I will debug this as soon as I get access to a few such
systems. -Dave Turner (turner@ameslab.gov)