Ready to test job_srun() (ref #194)

swift-lang · Oct 7, 2022 · 16df822 · 16df822
1 parent ffaab7d
commit 16df822
Show file tree

Hide file tree

Showing 3 changed files with 114 additions and 19 deletions.
diff --git a/turbine/code/export/job.swift b/turbine/code/export/job.swift
@@ -3,6 +3,6 @@
 // Various system-level configurations for application jobs.
 
 @par @dispatch=WORKER
-(int status) job_srun(int cores_per_job, int procs_per_job,
-                      string cmd_line[])
+(int status) job_srun(int cores_per_node, int cores_per_job, int procs_per_job,
+                      boolean bind, string cmd_line[])
   "turbine" "0.0" "job_srun_tcl";
diff --git a/turbine/code/lib/functions.tcl b/turbine/code/lib/functions.tcl
@@ -592,9 +592,37 @@ namespace eval turbine {
         }
         return $result
     }
+
+  proc contig { start count { step 1 } } {
+    set result [ list ]
+    set value $start
+    for { set i 0 } { $i < $count } { incr i } {
+      lappend result $value
+      incr value $step
+    }
+    return $result
+  }
+
+  # Break list L into count equal-size chunks (of size s)
+  proc fragment { L count } {
+    set result [ list ]
+    set n [ llength $L ]
+    set s [ expr $n / $count ]
+    set index 0
+    for { set c 0 } { $c < $count } { incr c } {
+      set chunk [ list ]
+      for { set i 0 } { $i < $s } { incr i } {
+        lappend chunk [ lindex $L [ expr $index + $i ] ]
+      }
+      lappend result $chunk
+      incr index $i
+    }
+
+    return $result
+  }
 }
 
 # Local Variables:
 # mode: tcl
-# tcl-indent-level: 4
+# tcl-indent-level: 2
 # End:
diff --git a/turbine/code/lib/job.tcl b/turbine/code/lib/job.tcl
@@ -6,17 +6,22 @@ namespace eval turbine {
 
   proc job_srun_tcl { outputs inputs } {
     set exit_code [ lindex $outputs 0 ]
-    set cpj [ lindex $inputs 0 ]
-    set ppj [ lindex $inputs 1 ]
-    set cmd_line [ lindex $inputs 2 ]
-    rule $inputs "turbine::job_srun_tcl_body $exit_code $cpj $ppj $cmd_line" \
+    set cpn  [ lindex $inputs 0 ]
+    set cpj  [ lindex $inputs 1 ]
+    set ppj  [ lindex $inputs 2 ]
+    set bind [ lindex $inputs 3 ]
+    set cmd_line [ lindex $inputs 4 ]
+    rule $inputs \
+        "turbine::job_srun_tcl_body $exit_code $cpn $cpj $ppj $bind $cmd_line" \
         type $turbine::WORK
   }
 
-  proc job_srun_tcl_body { exit_code cpj ppj cmd_line } {
+  proc job_srun_tcl_body { exit_code cpn cpj ppj bind cmd_line } {
     # Retrieve data (decr?)
-    set cpj_value [ retrieve_integer $cpj ]
-    set ppj_value [ retrieve_integer $ppj ]
+    set cpn_value  [ retrieve_integer $cpn ]
+    set cpj_value  [ retrieve_integer $cpj ]
+    set ppj_value  [ retrieve_integer $ppj ]
+    set bind_value [ retrieve_integer $bind ]
     # Unpack command line
     set D [ adlb::enumerate $cmd_line dict all 0 ]
     set cmd_value [ list ]
@@ -25,27 +30,89 @@ namespace eval turbine {
       lappend cmd_value [ dict get $D $k ]
     }
     # Run the user code
-    set exit_code_value [ job_srun_impl $cpj_value $ppj_value $cmd_value ]
+    set exit_code_value \
+        [ job_srun_impl $cpn_value $cpj_value $ppj_value $bind_value $cmd_value ]
     # Store result
     store_integer $exit_code $exit_code_value
   }
 
-  proc job_srun_impl { cpj ppj cmd } {
+  proc job_srun_impl { cpn cpj ppj bind cmd } {
+    # Setup and run the job.  Return a unix exit code.
+    global env
+    puts "turbine: srun: job_srun ..."
+
+    if $bind {
+      set cpu_bind [ bind_mask_cpu $cpn $cpj $ppj ]
+    } else {
+      set cpu_bind ""
+    }
+
+    puts "turbine: srun: job_srun -n $ppj -N 1 $cpu_bind $cmd"
+    puts "turbine: srun:   in PWD: $env(PWD)"
     try {
-      puts "turbine: srun: exec: srun -n $ppj $cmd"
-      set fp [ open "|srun -n $ppj $cmd" "r" ]
-      show fp
+      # Run the user job!  (with pipe to capture output)
+      set fp [ open "|srun -n $ppj -N 1 $cpu_bind $cmd 2>@1" "r" ]
       while { [ gets $fp line ] >= 0 } {
         puts "srun: $line"
       }
       close $fp
     } on error e {
-      puts "turbine: srun failed!"
-      puts "turbine: srun error message begin:"
-      puts $e
-      puts "turbine: srun error message end."
+      job_srun_error $e
       return 1
     }
     return 0
   }
+
+  proc bind_mask_cpu { cpn cpj ppj } {
+    # Set up the SLURM cpu binding
+    global env
+    set cpu_bind "--cpu-bind=verbose,mask_cpu:"
+    set offset $env(ADLB_RANK_OFFSET)
+    set ppn $env(PPN)
+    # puts "offset=$offset ppn=$ppn cpn=$cpn"
+    show offset ppn cpn cpj ppj
+
+    set L [ list ]
+    set start [ expr $offset * $cpj ]
+    set spacing [ expr $cpj / $ppj ]
+    set cpj_max [ expr $cpn / $ppn ]
+    show cpj_max
+    set start [ expr $cpj_max * $offset ]
+    # set S1 [ contig $start $cpj_max ]
+    # show S1
+    set step [ expr $cpj_max / $cpj ]
+    set S2 [ contig $start $cpj $step ]
+    show step S2
+    set K [ fragment $S2 $ppj ]
+    show K
+
+    # set cpu_ids [ join $L "," ]
+    # append cpu_bind $cpu_ids
+    set masks [ list ]
+    foreach chunk $K {
+      set mask [ list2mask $chunk ]
+      show mask
+      lappend masks $mask
+    }
+    show masks
+    append cpu_bind [ join $masks "," ]
+    return $cpu_bind
+  }
+
+  proc job_srun_error { e } {
+    puts "turbine: srun failed!"
+    puts "turbine: srun error message begin:"
+    puts $e
+    puts "turbine: srun error message end."
+  }
+
+  proc list2mask { L } {
+    set A 0
+    foreach i $L {
+      incr A [ expr 2 ** $i ]
+    }
+    puts $A
+    # printf "bitmap: %b" $A
+    return [ format "0x%X" $A ]
+  }
 }