diff --git a/bsc/garlic/control.nix b/bsc/garlic/control.nix
index 0e5daa8..268c928 100644
--- a/bsc/garlic/control.nix
+++ b/bsc/garlic/control.nix
@@ -14,7 +14,7 @@ stdenv.mkDerivation {
     mkdir -p $out/bin
     cat > $out/bin/run <<EOF
     #!/bin/sh
-    set -e
+    #set -e
     for n in {1..30}; do
       $program/bin/run
     done
diff --git a/bsc/garlic/default.nix b/bsc/garlic/default.nix
index 214b237..2821231 100644
--- a/bsc/garlic/default.nix
+++ b/bsc/garlic/default.nix
@@ -31,6 +31,14 @@ let
     controlWrapper = callPackage ./control.nix { };
     nixsetupWrapper = callPackage ./nix-setup.nix { };
     argvWrapper = callPackage ./argv.nix { };
+    statspyWrapper = callPackage ./statspy.nix { };
+    extraeWrapper = callPackage ./extrae.nix { };
+
+    # Perf is tied to a linux kernel specific version
+    linuxPackages = bsc.linuxPackages_4_4;
+    perfWrapper = callPackage ./perf.nix {
+      perf = linuxPackages.perf;
+    };
 
     exp = {
       nbody = {
diff --git a/bsc/garlic/exp/nbody/bs.nix b/bsc/garlic/exp/nbody/bs.nix
index 088d49a..0d99717 100644
--- a/bsc/garlic/exp/nbody/bs.nix
+++ b/bsc/garlic/exp/nbody/bs.nix
@@ -11,6 +11,9 @@
 , argvWrapper
 , controlWrapper
 , nixsetupWrapper
+, statspyWrapper
+, extraeWrapper
+, perfWrapper
 }:
 
 let
@@ -24,7 +27,7 @@ let
     gitBranch = "garlic/mpi+send";
     mpi = bsc.impi;
     particles = 1024*128;
-    timesteps = 10;
+    timesteps = 100;
     ntasksPerNode = "48";
     nodes = "1";
     time = "02:00:00";
@@ -46,6 +49,7 @@ let
   srun = app: srunWrapper {
     app = app;
     nixPrefix = "/gpfs/projects/bsc15/nix";
+    srunOptions = "--cpu-bind=verbose,rank";
   };
 
   argv = conf: app:
@@ -59,18 +63,47 @@ let
       argv = ''(-t ${toString timesteps} -p ${toString particles})'';
     };
 
+  statspy = app:
+    statspyWrapper {
+      app = app;
+    };
+
+  extrae = app:
+    extraeWrapper {
+      app = app;
+      traceLib = "mpi";
+      configFile = ./extrae.xml;
+    };
+
+  perf = app:
+    perfWrapper {
+      app = app;
+      perfArgs = "sched record -a";
+    };
+
   nbodyFn = conf:
     with conf;
     nbody.override { inherit cc mpi blocksize gitBranch; };
 
   pipeline = conf:
-    sbatch conf (
-      nixsetupWrapper (
-        controlWrapper (
+#   sbatch conf (
+#     nixsetupWrapper (
+#       controlWrapper (
           srun (
             nixsetupWrapper (
-              argv conf (
-                nbodyFn conf))))));
+#             extrae (
+#               perf (
+                  argv conf (
+                    nbodyFn conf
+                  )
+#               )
+#             )
+            )
+          )
+#       )
+#     )
+#   )
+  ;
 
   # Ideally it should look like this:
   #pipeline = sbatch nixsetup control argv nbodyFn;
diff --git a/bsc/garlic/exp/nbody/extrae.xml b/bsc/garlic/exp/nbody/extrae.xml
new file mode 100644
index 0000000..b9af29b
--- /dev/null
+++ b/bsc/garlic/exp/nbody/extrae.xml
@@ -0,0 +1,211 @@
+<?xml version='1.0'?>
+
+<!-- Here comes the Extrae configuration.
+     As a general rule, "enabled" means that the feature is enabled :)  If
+     it's not enabled, then the value can be set to some default.
+-->
+
+<!-- Must we activate the tracing? Which is the tracing mode? (detail/bursts) Where is it located? Which kind of trace? Version of the XML parser?-->
+<trace enabled="yes"
+ home="/nix/store/j80mlqa12d1baifg30jsx2smv90akzvc-extrae"
+ initial-mode="detail"
+ type="paraver"
+>
+
+  <!-- Configuration of some MPI dependant values -->
+  <mpi enabled="yes">
+    <!-- Gather counters in the MPI routines? -->
+    <counters enabled="yes" />
+  </mpi>
+
+  <!-- Emit information of the callstack -->
+  <callers enabled="yes">
+    <!-- At MPI calls, select depth level -->
+    <mpi enabled="yes">1-3</mpi>
+    <!-- At sampling points, select depth level -->
+		<sampling enabled="yes">1-5</sampling>
+		<!-- At dynamic memory system calls -->
+		<dynamic-memory enabled="no">1-3</dynamic-memory>
+		<!-- At I/O system calls -->
+		<input-output enabled="no">1-3</input-output>
+		<!-- At other system calls -->
+	  <syscall enabled="no">1-3</syscall>
+  </callers>
+
+  <!-- Configuration of some OpenMP dependant values -->
+  <openmp enabled="no" ompt="no">
+    <!-- If the library instruments OpenMP, shall we gather info about locks?
+         Obtaining such information can make the final trace quite large.
+    -->
+		<locks enabled="no" />
+		<!-- Gather info about taskloops? -->
+		<taskloop enabled="no" />
+    <!-- Gather counters in the OpenMP routines? -->
+    <counters enabled="yes" />
+  </openmp>
+
+  <!-- Configuration of some pthread dependant values -->
+  <pthread enabled="no">
+    <!-- If the library instruments pthread, shall we gather info about locks,
+	       mutexs and conds?
+         Obtaining such information can make the final trace quite large.
+    -->
+    <locks enabled="no" />
+    <!-- Gather counters in the pthread routines? -->
+    <counters enabled="yes" />
+  </pthread>
+
+  <!-- Configuration of User Functions -->
+  <user-functions enabled="no" list="/home/bsc41/bsc41273/user-functions.dat" exclude-automatic-functions="no">
+    <!-- Gather counters on the UF routines? -->
+    <counters enabled="yes" />
+  </user-functions>
+
+  <!-- Configure which software/hardware counters must be collected -->
+  <counters enabled="yes">
+    <!-- Configure the CPU hardware counters. You can define here as many sets
+         as you want. You can also define if MPI/OpenMP calls must report such
+         counters.
+         Starting-set property defines which set is chosen from every task.
+         Possible values are:
+           - cyclic : The sets are distributed in a cyclic fashion among all
+           tasks. So Task 0 takes set 1, Task 1 takes set 2,...
+           - block  : The sets are distributed in block fashion among all tasks.
+           Task [0..i-1] takes set 1, Task [i..2*i-1] takes set 2, ...
+           - Number : All the tasks will start with the given set 
+           (from 1..N).
+    -->
+    <cpu enabled="yes" starting-set-distribution="1">
+      <!-- In this example, we configure two sets of counters. The first will 
+           be changed into the second after 5 calls to some collective
+           operation on MPI_COMM_WORLD. Once the second is activated, it will
+           turn to the first after 5seconds (aprox. depending on the MPI calls
+           granularity)
+           If you want that any set be counting forever, just don't set
+           changeat-globalops, or, changeat-time.
+
+           Each set has it's own properties.
+           domain -> in which domain must PAPI obtain the information (see
+                       PAPI info)
+           changeat-globalops=num -> choose the next set after num
+                       MPI_COMM_WORLD operations
+           changeat-time=numTime -> choose the next set after num Time
+                       (for example 5s, 15m (for ms), 10M (for minutes),..)
+      -->
+      <set enabled="yes" domain="all">
+        PAPI_TOT_INS,PAPI_TOT_CYC
+      </set>
+    </cpu>
+
+    <!-- Do we want to gather information of the network counters?
+         Nowadays we can gather information about MX/GM cards.
+     -->
+    <network enabled="no" />
+
+    <!-- Obtain resource usage information -->
+    <resource-usage enabled="no" />
+
+    <!-- Obtain malloc statistics -->
+    <memory-usage enabled="no" />
+  </counters>
+
+  <!-- Define the characteristics of the tracing storage. If not defined,
+       or set, the tracing will send the traces to the current directory
+       with a default output name.
+  -->
+  <storage enabled="no">
+    <!-- The intermediate files will take the name of the application -->
+    <trace-prefix enabled="yes">TRACE</trace-prefix>
+    <!-- Stop the tracing when the intermediate file reaches this amount of MBs -->
+    <size enabled="no">5</size>
+    <!-- Where must we store the MPIT files while the app runs? -->
+    <temporal-directory enabled="yes">/scratch</temporal-directory>
+    <!-- Where must we store the MPIT files once the app ends? -->
+    <final-directory enabled="yes">/gpfs/scratch/bsc41/bsc41273</final-directory>
+  </storage>
+
+  <!-- Buffer configuration -->
+  <buffer enabled="yes">
+    <!-- How many events can we handle before any flush -->
+    <size enabled="yes">5000000</size>
+    <!-- Use the event buffer in a circular manner? You can use this option to
+         trace the last set of events. Needs MPI global routines operating on
+         MPI_COMM_WORLD communicator to be merged
+    -->
+    <circular enabled="no" />
+  </buffer>
+
+  <!-- Control tracing -->
+  <trace-control enabled="no">
+    <!-- We can start the application with a "latent tracing" and wake it up
+         once a control file is created. Use the property 'frequency' to
+         choose at which frequency this check must be done. If not supplied,
+         it will be checked every 100 global operations on MPI_COMM_WORLD.
+    -->
+    <file enabled="no" frequency="5M">/gpfs/scratch/bsc41/bsc41273/control</file>
+    <!-- 
+    -->
+    <global-ops enabled="no"></global-ops>
+  </trace-control>
+
+  <others enabled="yes">
+    <!-- Want to force a minimum amount of time of tracing? Here we force 10
+         minutes -->
+    <minimum-time enabled="no">10M</minimum-time>
+	<!-- Capture the following signals to finish cleanly -->
+    <finalize-on-signal enabled="yes" 
+      SIGUSR1="no" SIGUSR2="no" SIGINT="yes"
+      SIGQUIT="yes" SIGTERM="yes" SIGXCPU="yes"
+      SIGFPE="yes" SIGSEGV="yes" SIGABRT="yes"
+    />
+	<!-- Use instrumentation poitns to flush sampling buffer -->
+    <flush-sampling-buffer-at-instrumentation-point enabled="yes" />
+  </others>
+
+
+  <!-- Bursts library enabled? This requires an special library! -->
+  <bursts enabled="no">
+    <!-- Specify the threshold. This is mandatory! In this example, the
+         threshold is limitted to 500 microseconds
+     -->
+    <threshold enabled="yes">500u</threshold>
+    <!-- Report MPI statistics? -->
+    <mpi-statistics enabled="yes" />
+  </bursts>
+
+  <!-- Enable sampling capabilities using system clock.
+       Type may refer to: default, real, prof and virtual.
+       Period stands for the sampling period (50ms here)
+       plus a variability of 10ms, which means periods from 
+       45 to 55ms.
+  -->
+  <sampling enabled="no" type="default" period="50m" variability="10m" />
+
+  <!-- Enable dynamic memory instrumentation (experimental) -->
+  <dynamic-memory enabled="no" />
+
+  <!-- Enable I/O (read, write) instrumentation (experimental) -->
+  <input-output enabled="no" internals="no"/>
+
+	<!-- Enable system calls instrumentation (experimental) -->
+  <syscall enabled="no" />
+
+  <!-- Do merge the intermediate tracefiles into the final tracefile?
+       Named according to the binary name
+       options: 
+       synchronization = { default, task, node, no } (default is node)
+       max-memory = Number (in Mbytes) max memory used in merge step
+       joint-states = { yes, no } generate joint states?
+       keep-mpits = { yes, no } keep mpit files after merge?
+  -->
+  <merge enabled="yes" 
+    synchronization="default"
+    tree-fan-out="16"
+    max-memory="512"
+    joint-states="yes"
+    keep-mpits="yes"
+    sort-addresses="yes"
+    overwrite="yes"
+  />
+
+</trace>
diff --git a/bsc/garlic/exp/osu/latency.nix b/bsc/garlic/exp/osu/latency.nix
index 3a3a777..77970c6 100644
--- a/bsc/garlic/exp/osu/latency.nix
+++ b/bsc/garlic/exp/osu/latency.nix
@@ -64,7 +64,17 @@ let
     with conf;
     bsc.osumb.override { inherit mpi; };
 
-  pipeline = conf: sbatch conf (srun (nixsetupWrapper (argv (osumbFn conf))));
+
+  pipeline = conf:
+    sbatch conf (
+      nixsetupWrapper (
+        controlWrapper (
+          srun (
+            nixsetupWrapper (
+              argv (
+                osumbFn conf))))));
+
+  #pipeline = conf: sbatch conf (srun (nixsetupWrapper (argv (osumbFn conf))));
   #pipeline = conf: sbatch conf (srun (nixsetupWrapper (argv bsc.osumb)));
 
   # Ideally it should look like this:
diff --git a/bsc/garlic/extrae.nix b/bsc/garlic/extrae.nix
new file mode 100644
index 0000000..174636f
--- /dev/null
+++ b/bsc/garlic/extrae.nix
@@ -0,0 +1,30 @@
+{
+  stdenv
+, bash
+, extrae
+}:
+
+{
+  app
+, traceLib ? "mpi"
+, configFile
+, program ? "bin/run"
+}:
+
+stdenv.mkDerivation {
+  name = "${app.name}-extrae";
+  preferLocalBuild = true;
+  phases = [ "installPhase" ];
+  installPhase = ''
+    mkdir -p $out/bin
+    cat > $out/bin/run <<EOF
+    #!${bash}/bin/bash
+
+    export EXTRAE_HOME=${extrae}
+    export LD_PRELOAD=${extrae}/lib/lib${traceLib}trace.so:$LD_PRELOAD
+    export EXTRAE_CONFIG_FILE=${configFile}
+    exec ${app}/${program}
+    EOF
+    chmod +x $out/bin/run
+  '';
+}
diff --git a/bsc/garlic/nbody/default.nix b/bsc/garlic/nbody/default.nix
index 3de6e2e..4b88ee1 100644
--- a/bsc/garlic/nbody/default.nix
+++ b/bsc/garlic/nbody/default.nix
@@ -12,10 +12,12 @@ with stdenv.lib;
 stdenv.mkDerivation rec {
   name = "nbody";
 
-  src = builtins.fetchGit {
-    url = "${gitURL}";
-    ref = "${gitBranch}";
-  };
+  src = /home/Computational/rarias/bsc-nixpkgs/manual/nbody;
+
+  #src = builtins.fetchGit {
+  #  url = "${gitURL}";
+  #  ref = "${gitBranch}";
+  #};
 
   buildInputs = [
     cc
diff --git a/bsc/garlic/perf.nix b/bsc/garlic/perf.nix
new file mode 100644
index 0000000..5117f0a
--- /dev/null
+++ b/bsc/garlic/perf.nix
@@ -0,0 +1,26 @@
+{
+  stdenv
+, bash
+, perf
+}:
+
+{
+  app
+, perfArgs ? "record -a"
+, program ? "bin/run"
+}:
+
+stdenv.mkDerivation {
+  name = "${app.name}-perf";
+  preferLocalBuild = true;
+  phases = [ "installPhase" ];
+  installPhase = ''
+    mkdir -p $out/bin
+    cat > $out/bin/run <<EOF
+    #!${bash}/bin/bash
+
+    exec ${perf}/bin/perf ${perfArgs} ${app}/${program}
+    EOF
+    chmod +x $out/bin/run
+  '';
+}
diff --git a/bsc/garlic/sbatch.nix b/bsc/garlic/sbatch.nix
index 41bffb1..48c2857 100644
--- a/bsc/garlic/sbatch.nix
+++ b/bsc/garlic/sbatch.nix
@@ -39,6 +39,7 @@ stdenv.mkDerivation rec {
   src = ./.;
 
   buildInputs = [ app ];
+  phases = [ "installPhase" ];
 
   #SBATCH --tasks-per-node=48
   #SBATCH --ntasks-per-socket=24
diff --git a/bsc/garlic/srun.nix b/bsc/garlic/srun.nix
index 3ede737..e304fc7 100644
--- a/bsc/garlic/srun.nix
+++ b/bsc/garlic/srun.nix
@@ -1,7 +1,11 @@
 {
   stdenv
 }:
-{ app , nixPrefix ? "" }:
+{
+  app
+, nixPrefix ? ""
+, srunOptions ? ""
+}:
 
 stdenv.mkDerivation rec {
   name = "${app.name}-srun";
@@ -13,7 +17,7 @@ stdenv.mkDerivation rec {
     mkdir -p $out/bin
     cat > $out/bin/run <<EOF
     #!/bin/sh
-    exec srun --mpi=pmi2 ${nixPrefix}${app}/bin/run
+    exec srun --mpi=pmi2 ${srunOptions} ${nixPrefix}${app}/bin/run
     EOF
     chmod +x $out/bin/run
   '';
diff --git a/bsc/garlic/statspy.nix b/bsc/garlic/statspy.nix
new file mode 100644
index 0000000..d039353
--- /dev/null
+++ b/bsc/garlic/statspy.nix
@@ -0,0 +1,29 @@
+{
+  stdenv
+, bash
+}:
+
+{
+  app
+, outputDir ? "."
+, program ? "bin/run"
+}:
+
+stdenv.mkDerivation {
+  name = "${app.name}-statspy";
+  preferLocalBuild = true;
+  phases = [ "installPhase" ];
+  installPhase = ''
+    mkdir -p $out/bin
+    cat > $out/bin/run <<EOF
+    #!${bash}/bin/bash
+
+    mkdir -p ${outputDir}
+    cat /proc/[0-9]*/stat | sort -n > ${outputDir}/statspy.\$(date +%s.%3N).begin
+    ${app}/${program}
+    cat /proc/[0-9]*/stat | sort -n > ${outputDir}/statspy.\$(date +%s.%3N).end
+
+    EOF
+    chmod +x $out/bin/run
+  '';
+}