f518e09eab396bbdbaa49ff8420307d04c6e88a3
hosttrace: application tracing with static/dynamic binary instrumentation and perfetto
1. load necessary modules
module load julia
module load dyninst
module load rocm/VERSION
2. to install julia packages, only need it the first time.
$ julia
type ']' to get into pkg mode, then
(@v1.6) pkg> add JSON, DataFrames, Dates, CSV, Chain, PrettyTables
3. hosttrace usage
export PATH=$PATH:$HTRACE_PATH/bin
hosttrace --help
4. instrument the binaries
hosttrace -L $HTRACE_PATH/bin/libhosttrace.so -o app.inst -- path_to_your_app
hosttrace -L $HTRACE_PATH/bin/libhosttrace.so -E 'hipApiName|hipGetCmdName' -o libamdhip64.so.4 -- /opt/rocm-VERSION/lib/libamdhip64.so.4
hosttrace -L $HTRACE_PATH/bin/libhosttrace.so -E 'rocr::atomic|rocr::core|rocr::HSA' -o libhsa-runtime64.so.1 -- /opt/rocm-VERSION/lib/libhsa-runtime64.so.1
5. run the app
export LD_LIBRARY_PATH=.:$LD_LIBRARY_PATH
rocprof --hip-trace --roctx-trace --stats ./app.inst
6. you may need to increase the buffer size if the hosttrace.perfetto-trace file is close to 1GB. for example to set it to 10GB.
export HOSTTRACE_BUFFER_SIZE_KB=10240000
7. merge the traces from rocprof and hosttrace
julia $HTRACE_PATH/bin/merge-trace.jl results.json hosttrace.perfetto-trace*
8. another mode of Perfetto tracing is to use system backend. To do it:
in a separate window
pkill traced; traced --background; perfetto --out ./htrace.out --txt -c $HTRACE/roctrace.cfg
then in the app running window do this before running rocprof or other app
export HOSTTRACE_BACKEND_SYSTEM=1
for the merge use the htrace.out
julia $HTRACE_PATH/bin/merge-trace.jl results.json htrace.out
Description
Languages
C++
67.5%
C
20.6%
Python
6.6%
CMake
3.4%
Shell
0.6%
Other
1.1%