Mentions légales du service

Skip to content
Snippets Groups Projects
Unverified Commit 39d30830 authored by Christian Heinrich's avatar Christian Heinrich
Browse files

[Computation] Added computation-calibration.org 

This script allows users to account for caches. The output is
an input file for SMPI, see the smpi/trace-call-location option (and friends).
parent b9b3fbe9
No related branches found
No related tags found
No related merge requests found
Hide whitespace changes
Inline Side-by-side
src/computation-calibration/computation-calibration.org 0 → 100644
+ 292
− 0
View file @ 39d30830
* Set some R options...
#+begin_src R :results output :session *R* :exports both
options(width=200)
library(plyr)
library(ggplot2)
#+end_src
#+RESULTS:
** Remove unneccessary information from the traces
#+begin_src bash :results output :exports both :session foo
cat /tmp/SMPI_LU_CLASS=C_PROCS=12_FREQ=2300_CLUSTER=TAURUS_trace_dump | grep State, | grep MPI_ | \
sed -e 's/()//' -e 's/MPI_STATE, //ig' -e 's/State, //ig' -e 's/rank-//' -e 's/PMPI_/MPI_/' | grep MPI_ | tr 'A-Z' 'a-z' > /tmp/smpi_12.csv
cat /tmp/MPI_LU_C_NODES=1_CORES=12_NPROCS=12_FREQUENCY=2300_CLUSTER=taurus.csv | grep State, | \
sed -e 's/action_/MPI_/' -e 's/rank//' -e 's/ MPI_STATE,//' -e 's/STATE, //ig' -e 's/()//' | \
sed -E 's/_([[:lower:]])/_\U\1/' | grep MPI_ | tr 'A-Z' 'a-z' > /tmp/mpi_12.csv
#+end_src
#+RESULTS:
* Load data in R
#+begin_src R :results output :session *R* :exports both
df_mpi12 = read.csv("/tmp/mpi_12.csv", header=F, strip.white=T);
names(df_mpi12) = c("rank", "start", "end", "duration", "level", "state");
df_smpi12 = read.csv("/tmp/smpi_12.csv", header=F, strip.white=T);
names(df_smpi12) = c("rank", "start", "end", "duration", "level", "state", "Filename", "Linenumber");
df_mpi12$origin = as.factor("MPI 12 cores")
df_smpi12$origin = as.factor("SMPI 12 cores")
#+end_src
#+RESULTS:
#+begin_src R :results output :session *R* :exports both
str(df_mpi12)
str(df_smpi12)
#+end_src
#+RESULTS:
#+begin_example
'data.frame': 2757182 obs. of 7 variables:
$ rank : int 9 9 9 9 9 9 9 9 9 9 ...
$ start : num 0.0236 0.0906 0.0907 0.0914 0.0918 ...
$ end : num 0.0812 0.0906 0.0907 0.0918 0.0918 ...
$ duration: num 0.057596 0.000007 0.000001 0.000346 0.00002 ...
$ level : num 4 4 4 5 5 5 5 5 5 5 ...
$ state : Factor w/ 11 levels "mpi_allreduce",..: 7 4 5 3 3 3 3 3 3 3 ...
$ origin : Factor w/ 1 level "MPI 12 cores": 1 1 1 1 1 1 1 1 1 1 ...
'data.frame': 2757156 obs. of 9 variables:
$ rank : int 8 8 8 8 8 8 8 8 8 8 ...
$ start : num 0 0.101 0.101 0.101 0.101 ...
$ end : num 0 0.101 0.101 0.101 0.101 ...
$ duration : num 0.0 4.3e-05 1.5e-05 1.5e-05 1.5e-05 1.6e-05 1.6e-05 1.5e-05 1.6e-05 1.6e-05 ...
$ level : num 1 1 1 1 1 1 1 1 1 1 ...
$ state : Factor w/ 9 levels "mpi_allreduce",..: 5 3 3 3 3 3 3 3 3 3 ...
$ Filename : Factor w/ 11 levels "bcast_inputs.f",..: 8 1 1 1 1 1 1 1 1 1 ...
$ Linenumber: int 31 28 29 30 31 32 33 34 35 36 ...
$ origin : Factor w/ 1 level "SMPI 12 cores": 1 1 1 1 1 1 1 1 1 1 ...
#+end_example
#+begin_src R :results output :session *R* :exports both
unique(df_mpi12$state)
unique(df_smpi12$state)
#+end_src
#+RESULTS:
: [1] mpi_init mpi_comm_rank mpi_comm_size mpi_bcast mpi_irecv mpi_send mpi_wait mpi_allreduce mpi_barrier mpi_recv mpi_finalize
: Levels: mpi_allreduce mpi_barrier mpi_bcast mpi_comm_rank mpi_comm_size mpi_finalize mpi_init mpi_irecv mpi_recv mpi_send mpi_wait
: [1] mpi_init mpi_bcast mpi_send mpi_irecv mpi_wait mpi_allreduce mpi_barrier mpi_recv mpi_finalize
: Levels: mpi_allreduce mpi_barrier mpi_bcast mpi_finalize mpi_init mpi_irecv mpi_recv mpi_send mpi_wait
Let's get rid of =mpi_comm_size= and =mpi_comm_rank=:
#+begin_src R :results output :session *R* :exports both
df_mpi12 = df_mpi12[!(df_mpi12$state %in% c("mpi_comm_size","mpi_comm_rank")),]
#+end_src
#+RESULTS:
We just removed the rows that had those two values in the column
state; however, since the column =state= is a factor, they will still
show up. (=factor= means that there is a set of values that are possible)
#+begin_src R :results output :session *R* :exports both
unique(df_mpi12$state)
#+end_src
#+RESULTS:
: [1] mpi_init mpi_bcast mpi_irecv mpi_send mpi_wait mpi_allreduce mpi_barrier mpi_recv mpi_finalize
: Levels: mpi_allreduce mpi_barrier mpi_bcast mpi_comm_rank mpi_comm_size mpi_finalize mpi_init mpi_irecv mpi_recv mpi_send mpi_wait
Let's verify it's coherent with SMPI:
#+begin_src R :results output :session *R* :exports both
length(unique(df_mpi12$state)) == length(unique(df_smpi12$state))
#+end_src
#+RESULTS:
: [1] TRUE
#+begin_src R :results output :session *R* :exports both
df_mpi12 = df_mpi12[names(df_mpi12)!="level"];
df_smpi12 = df_smpi12[names(df_smpi12)!="level"];
#+end_src
#+RESULTS:
Let's compare the amount of events:
#+begin_src R :results output :session *R* :exports both
dim(df_mpi12)
dim(df_smpi12)
#+end_src
#+RESULTS:
: [1] 2757157 6
: [1] 2757156 8
* Fix missing MPI_Finalize call (bug in paje)
Let's add the missing MPI_Finalize to rank 0. This is a bug in PAJE
that should have been fixed in recent versions (so the following is
not necessary any longer).
#+begin_src R :results output :session *R* :exports both
df_smpi12_fake = head(df_smpi12[df_smpi12$state=="mpi_finalize" & df_smpi12$end == max(df_smpi12$end),],n=1)
df_smpi12_fake$rank=0
df_smpi12_fake$duration=df_smpi12_fake$end-df_smpi12_fake$start
df_smpi12 = rbind(df_smpi12, df_smpi12_fake)
#+end_src
#+RESULTS:
The MPI dataframe is lacking filename and linenumber, so let's
just use the ones from SMPI (they are from the original source code
files, so they are appropriate for MPI as well).
Let's sort by column rank first:
#+begin_src R :results output :session *R* :exports both
df_smpi12 = df_smpi12[with(df_smpi12,order(rank,start)),]; # Sort df_smpi
df_mpi12 = df_mpi12[with(df_mpi12,order(rank,start)),];
df_mpi12$Filename = df_smpi12$Filename
df_mpi12$Linenumber = df_smpi12$Linenumber
#+end_src
#+RESULTS:
* Compute the =computation= timings!
#+BEGIN_SRC R :results output :session *R* :exports both
duration_compute = function(df) {
ndf = data.frame();
df = df[with(df,order(rank,start)),];
origin = unique(df$origin)
for(i in (sort(unique(df$rank)))) {
start = df[df$rank==i,]$start;
end = df[df$rank==i,]$end;
l = length(end);
end = c(0,end[1:(l-1)]); # Computation starts at time 0
startline = c(0, df[df$rank==i,]$Linenumber[1:(l-1)]);
startfile = c("", as.character(df[df$rank==i,]$Filename[1:(l-1)]));
endline = df[df$rank==i,]$Linenumber;
endfile = df[df$rank==i,]$Filename;
ndf = rbind(ndf, data.frame(rank=i, start=end, end=start,
duration=start-end, state="Computing", origin=origin,
startline=startline, startfile=startfile, endline=endline,
endfile=endfile));
}
ndf$idx = 1:length(ndf$duration)
ndf;
}
df_mpi12_comp = duration_compute(df_mpi12);
df_smpi12_comp = duration_compute(df_smpi12);
#+END_SRC
#+RESULTS:
* Generate speedup file
#+begin_src R :results output :session *R* :exports both
ddf_mpi=ddply(df_mpi12_comp,c("startfile","startline","endfile","endline"),summarize,count=length(duration),duration=sum(duration));
ddf_mpi$origin="MPI12"
ddf_mpi = ddf_mpi[with(ddf_mpi,order(startfile,startline,endfile,endline)),];
ddf_smpi=ddply(df_smpi12_comp,c("startfile","startline","endfile","endline"),summarize,count=length(duration),duration=sum(duration));
ddf_smpi$origin="SMPI"
ddf_smpi = ddf_smpi[with(ddf_smpi,order(startfile,startline,endfile,endline)),];
ddf=ddf_smpi
ddf$rlduration = ddf_mpi$duration
ddf=ddf[!names(ddf)=="origin"]
ddf = ddf[ddf$duration>1,]
head(ddf)
#+end_src
#+RESULTS:
: startfile startline endfile endline count duration rlduration
: 11 bcast_inputs.f 37 exchange_3.f 42 9 1.897406 11.459090
: 15 exchange_1.f 30 exchange_1.f 48 240960 6.814807 0.464526
: 16 exchange_1.f 30 exchange_1.f 113 80320 47.611578 36.714682
: 17 exchange_1.f 30 exchange_1.f 130 40160 23.014585 17.711225
: 18 exchange_1.f 48 exchange_1.f 30 39909 40.914132 31.306660
: 19 exchange_1.f 48 exchange_1.f 113 240960 140.137475 110.589122
#+begin_src R :results output :session *R* :exports both
ddf$ratio = ddf$duration / ddf$rlduration
ddf
sum(ddf$duration)
sum(ddf$rlduration)
#+end_src
#+RESULTS:
#+begin_example
startfile startline endfile endline count duration rlduration ratio
11 bcast_inputs.f 37 exchange_3.f 42 9 1.897406 11.459090 0.1655809
15 exchange_1.f 30 exchange_1.f 48 240960 6.814807 0.464526 14.6704533
16 exchange_1.f 30 exchange_1.f 113 80320 47.611578 36.714682 1.2967994
17 exchange_1.f 30 exchange_1.f 130 40160 23.014585 17.711225 1.2994350
18 exchange_1.f 48 exchange_1.f 30 39909 40.914132 31.306660 1.3068827
19 exchange_1.f 48 exchange_1.f 113 240960 140.137475 110.589122 1.2671904
20 exchange_1.f 48 exchange_1.f 130 40160 23.105738 18.117715 1.2753119
22 exchange_1.f 68 exchange_1.f 86 240960 7.078204 0.496461 14.2573213
23 exchange_1.f 68 exchange_1.f 149 80320 49.818296 36.917500 1.3494493
24 exchange_1.f 68 exchange_1.f 166 40160 24.672909 18.251679 1.3518159
25 exchange_1.f 86 exchange_1.f 68 39909 43.393137 32.066507 1.3532231
26 exchange_1.f 86 exchange_1.f 149 240960 148.546141 110.436460 1.3450824
27 exchange_1.f 86 exchange_1.f 166 40160 25.060315 18.573204 1.3492726
28 exchange_1.f 86 exchange_3.f 73 250 3.062204 2.915671 1.0502570
30 exchange_1.f 113 exchange_1.f 30 79818 55.237256 28.561952 1.9339454
31 exchange_1.f 113 exchange_1.f 48 39909 27.860205 14.109157 1.9746187
33 exchange_1.f 113 exchange_1.f 130 240960 3.661200 0.312652 11.7101442
34 exchange_1.f 130 exchange_1.f 30 239454 168.361242 85.475785 1.9696952
35 exchange_1.f 130 exchange_1.f 48 39909 28.245751 14.382586 1.9638854
36 exchange_1.f 130 exchange_1.f 68 1506 1.002664 0.524498 1.9116641
38 exchange_1.f 130 exchange_1.f 113 39909 48.107018 32.180717 1.4949020
39 exchange_1.f 149 exchange_1.f 68 79818 55.569462 28.087507 1.9784405
40 exchange_1.f 149 exchange_1.f 86 39909 27.010726 13.315344 2.0285414
41 exchange_1.f 149 exchange_1.f 166 240960 3.535476 0.322482 10.9633282
42 exchange_1.f 149 exchange_3.f 42 750 7.899729 8.361331 0.9447932
44 exchange_1.f 166 exchange_1.f 68 239454 164.925555 83.439602 1.9765861
45 exchange_1.f 166 exchange_1.f 86 39909 27.118416 13.617528 1.9914346
46 exchange_1.f 166 exchange_1.f 149 39909 45.719675 30.848295 1.4820811
47 exchange_1.f 166 exchange_3.f 42 1500 15.617569 16.850112 0.9268525
48 exchange_1.f 166 exchange_3.f 73 500 5.500635 6.143575 0.8953476
50 exchange_3.f 42 exchange_3.f 73 1524 1.710175 1.078730 1.5853596
54 exchange_3.f 86 exchange_3.f 108 1524 3.988559 0.487534 8.1810889
55 exchange_3.f 86 exchange_3.f 139 762 2.413731 0.936752 2.5767023
56 exchange_3.f 108 exchange_3.f 139 1524 1.638622 0.771228 2.1246920
59 exchange_3.f 139 exchange_3.f 178 508 15.912982 13.404639 1.1871250
60 exchange_3.f 139 exchange_3.f 209 254 7.969534 6.665093 1.1957123
61 exchange_3.f 152 exchange_3.f 178 1524 53.239456 42.864135 1.2420513
62 exchange_3.f 152 exchange_3.f 209 762 26.780502 21.316899 1.2563038
71 exchange_3.f 275 exchange_1.f 30 747 49.679879 55.816729 0.8900536
72 exchange_3.f 275 exchange_1.f 48 249 16.580075 18.731022 0.8851666
78 exchange_3.f 288 exchange_1.f 30 1494 101.176528 113.250811 0.8933846
79 exchange_3.f 288 exchange_1.f 48 249 17.034232 19.059389 0.8937449
80 exchange_3.f 288 exchange_1.f 113 249 16.926680 18.798087 0.9004469
83 exchange_3.f 288 l2norm.f 57 24 1.613920 1.784543 0.9043884
103 init_comm.f 31 bcast_inputs.f 28 12 1.215429 0.123510 9.8407335
119 ssor.f 210 exchange_3.f 42 9 1.651463 11.049886 0.1494552
[1] 1590.031
[1] 1148.693
#+end_example
Let's write this to a file so that we can import it to SMPI
#+begin_src R :results output :exports both :width 600 :height 400 :session *R*
tmp_df=data.frame("start_stop"=ddf$startfile,"ratio"=ddf$ratio)
tmp_df$"start_stop"=paste(ddf$startfile, ddf$startline, ddf$endfile, ddf$endline, sep=":")
write.csv(tmp_df, file="/tmp/2016-08-11-smpi-mpi-ratio.csv", row.names=FALSE)
head(tmp_df)
rm(tmp_df)
#+end_src
#+RESULTS:
: start_stop ratio
: 1 bcast_inputs.f:37:exchange_3.f:42 0.1655809
: 2 exchange_1.f:30:exchange_1.f:48 14.6704533
: 3 exchange_1.f:30:exchange_1.f:113 1.2967994
: 4 exchange_1.f:30:exchange_1.f:130 1.2994350
: 5 exchange_1.f:48:exchange_1.f:30 1.3068827
: 6 exchange_1.f:48:exchange_1.f:113 1.2671904
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment