XiTAO RUNTIME

XiTAO is a lightweight layer built on top of modern C++ features with the goals of being low-overhead and serving as a development platform for testing scheduling and resource management algorithms. XiTAO is built on a generalized model of task which that assembles (1) concurrency, (2) an embedded scheduler and (3) a resizeable resource container. These TAOs (Task Assembly Objects) are moldable entities that can be scheduled into elastic resource partitioning, aka "elastic places". XiTAO targets better mapping between tasks and hardware resources such as cores, caches or interconnect bandwidth. Therefore, among other features, XiTAO provides fast parallelism at low overhead, with constructive sharing and interference-avoidance. Recently, XiTAO has been extended with novel high-level programming constructs, and with instrospective scheduling techniques that allow it to adapt to dynamic events as well as better target the goals of energy-efficiency.

The following figure shows the simplified architecture of the XiTAO runtime

Building the XiTAO Library

make lib

Running Dot Product Example

cd benchmarks/dotproduct
make dotprod
./dotprod 8192 2 16

Available command line options for the XiTAO runtime

By initializing the command line arguments to the runtime using xitao_init(argc, argv), run your DAG application with ./bin --xitao_args="-h" to get the available options, such as:

Usage: --xitao_args= [options]
Long option (short option)               : Description (Default value)
 --wstealing (-w) [0/1]                  : Enable/Disable work-stealing (1)
 --perfmodel (-p) [0/1]                  : Enable/Disable performance modeling  (1)
 --nthreads (-t)                         : The number of worker threads (8)
 --idletries (-i)                        : The number of idle tries before a steal attempt (100)
 --minparcost (-c) [0/1]                 : model 1 (parallel cost) - 0 (parallel time) (1)
 --oldtickweight (-o)                    : Weight of old tick versus new tick (4)
 --refreshtablefreq (-t)                 : How often to attempt a random moldability to heat the table (10)
 --mold (-m)                             : Enable/Disable dynamic moldability (1)
 --help (-h)                             : Show this help document

Explanation for the Dot Product Example

read documentation.pdf

Running Synthetic DAGs Benchmark

cd benchmarks/syntheticDAGs
make clean; make run
./synbench <Block Side Length> <Resource Width> <TAO Mul Count> <TAO Copy Count> <TAO Stencil Count> <Degree of Parallelism>

Turning on Debug Trace

cd ./include/
vim config.h
#define DEBUG

Changing PTT Hardware Topology

Create a file with number of lines = number of threads + 1

-- first line contains affinity for each thread

-- the subsequent lines correspond to the threads listed in order, in which the thread is a leader. The line contains the allowed resource partition widths

-- set environment variable XITAO_LAYOUT_PATH to the absolute path of the layout file

Example: suppose you have 6 threads, and you would like to allow widths 1 and 2 where appropriate. A possible file (assuming contiguous core affinity ids)

0 1 2 3 4 5
1 2
1
1 2
1
1 2
1

ACKNOWLEDGEMENT

This work has been supported by EU H2020 ICT project LEGaTO, contract #780681.