rbbt-util 5.28.14 → 5.30.0

data/lib/rbbt/workflow/step/dependencies.rb

@@ -103,7 +103,7 @@ class Step
       end
 
       job.dup_inputs unless status == 'done' or job.started?
-      job.init_info(status == 'noinfo') unless status == 'waiting' || status == 'done' || job.started?
+      job.init_info(status == 'noinfo') unless status == 'waiting' || status == 'done' || job.started? || ! Workflow.job_path?(job.path)
 
       canfail = ComputeDependency === job && job.canfail?
     end
@@ -130,7 +130,6 @@ class Step
     (inputs.flatten.select{|i| Step === i} + inputs.flatten.select{|dep| Path === dep && Step === dep.resource}.collect{|dep| dep.resource})
   end
 
-
   def execute_dependency(dependency, log = true)
     task_name = self.task_name
     canfail_paths = self.canfail_paths

data/lib/rbbt/workflow/step/run.rb

@@ -122,7 +122,6 @@ class Step
       reject{|dependency| (defined?(WorkflowRemoteClient) && WorkflowRemoteClient::RemoteStep === dependency) || Open.remote?(dependency.path) }.
       reject{|dependency| dependency.error? }.
       #select{|dependency| Open.exists?(dependency.path) || ((Open.exists?(dependency.info_file) && (dependency.status == :cleaned) || dependency.status == :waiting)) }.
-      #select{|dependency| Open.exists?(dependency.path) || ((Open.exists?(dependency.info_file) && (dependency.status == :cleaned) || dependency.status == :waiting)) }.
       select{|dependency| dependency.updatable? }.
       collect{|dependency| Workflow.relocate_dependency(self, dependency)}
   end
@@ -374,7 +373,6 @@ class Step
                 Log.exception $!
               ensure
                 Step.purge_stream_cache
-                set_info :pid, nil
                 Open.rm pid_file if Open.exist?(pid_file)
               end
             end
@@ -389,7 +387,6 @@ class Step
                 _clean_finished
               rescue
                 stop_dependencies
-                set_info :pid, nil
                 Open.rm pid_file if Open.exist?(pid_file)
               end
             end
@@ -450,7 +447,7 @@ class Step
     ensure 
       no_load = false unless IO === result
       Open.rm pid_file if Open.exist?(pid_file) unless no_load
-      set_info :pid, nil unless no_load
+      #set_info :pid, nil unless no_load
     end
   end
 
@@ -560,7 +557,7 @@ class Step
           RbbtSemaphore.post_semaphore(semaphore) if semaphore
           Kernel.exit! -1
         end
-        set_info :pid, nil
+        #set_info :pid, nil
       ensure
         RbbtSemaphore.post_semaphore(semaphore) if semaphore
       end

data/lib/rbbt/workflow/usage.rb

@@ -57,7 +57,7 @@ module Task
       puts Log.color(:magenta, "Input select options")
       puts
       selects.collect{|p| p}.uniq.each do |input,options|
-        puts Log.color(:blue, input.to_s + ": ") << Misc.format_paragraph(options.collect{|o| o.to_s} * ", ") << "\n"
+        puts Log.color(:blue, input.to_s + ": ") << Misc.format_paragraph(options.collect{|o| Array === o ? o.first.to_s : o.to_s} * ", ") << "\n"
         puts unless Log.compact
       end
       puts

data/lib/rbbt/workflow/util/orchestrator.rb

@@ -26,15 +26,25 @@ module Workflow
       workload
     end
 
+    def self.workload(jobs)
+      jobs.inject({}) do |acc,job| 
+          Orchestrator.job_workload(job).each do |j,d|
+            acc[j] = d unless acc.keys.collect{|k| k.path }.include? j.path
+          end
+          acc
+        end
+    end
+
     def self.job_rules(rules, job)
       workflow = job.workflow.to_s
       task_name = job.task_name.to_s
+      defaults = rules["defaults"] || {}
 
-      return IndiferentHash.setup(rules["defaults"]) unless rules[workflow]
-      return IndiferentHash.setup(rules["defaults"]) unless rules[workflow][task_name]
+      return IndiferentHash.setup(defaults) unless rules[workflow]
+      return IndiferentHash.setup(defaults) unless rules[workflow][task_name]
 
       job_rules = IndiferentHash.setup(rules[workflow][task_name])
-      rules["defaults"].each{|k,v| job_rules[k] = v if job_rules[k].nil? } if rules["defaults"]
+      defaults.each{|k,v| job_rules[k] = v if job_rules[k].nil? } if defaults
       job_rules
     end
 
@@ -169,12 +179,7 @@ module Workflow
     def process(rules, jobs)
       begin
 
-        workload = jobs.inject({}) do |acc,job| 
-          Orchestrator.job_workload(job).each do |j,d|
-            acc[j] = d unless acc.keys.collect{|k| k.path }.include? j.path
-          end
-          acc
-        end
+        workload = Orchestrator.workload(jobs)
         all_jobs = workload.keys
 
         top_level_jobs = jobs.collect{|job| job.path }

data/lib/rbbt/workflow/util/provenance.rb

@@ -22,11 +22,14 @@ class Step
   end
 
   def self.prov_report_msg(status, name, path, info = nil)
-    parts = path.sub(/\{.*/,'').sub(/#{Regexp.quote(name)}$/,'').split "/"
+    parts = path.sub(/\{.*/,'').split "/"
 
+    parts.pop
+    
     task = Log.color(:yellow, parts.pop)
     workflow = Log.color(:magenta, parts.pop)
-    if status.to_s == 'noinfo' and parts.last != 'jobs'
+    #if status.to_s == 'noinfo' && parts.last != 'jobs'
+    if ! Workflow.job_path?(path)
       task, status, workflow = Log.color(:yellow, info[:task_name]), Log.color(:green, "file"), Log.color(:magenta, "-")
     end
 

data/share/rbbt_commands/slurm/clean

@@ -0,0 +1,165 @@
+#!/usr/bin/env ruby
+
+require 'rbbt-util'
+require 'rbbt/util/simpleopt'
+
+#$0 = "rbbt #{$previous_commands*""} #{ File.basename(__FILE__) }" if $previous_commands
+
+options = SOPT.setup <<EOF
+
+Clean error or aborted jobs
+
+$ rbbt mnl [options] 
+
+-h--help Print this help
+-d--done Done jobs only
+-e--error Error jobs only
+-a--aborted SLURM aboted jobs
+-j--job* Job ids
+-s--search* Regular expression
+-t--tail* Show the last lines of the STDERR
+-SBP--sbatch_parameters show sbatch parameters
+-dr--dry_run Do not erase anything
+EOF
+
+if options[:help]
+ if defined? rbbt_usage
+  rbbt_usage 
+ else
+  puts SOPT.doc
+ end
+ exit 0
+end
+
+Log.severity = 4
+done, error, aborted, jobid, search, tail, sbatch_parameters, dry_run = options.values_at :done, :error, :aborted, :job, :search, :tail, :sbatch_parameters, :dry_run
+
+workdir = File.expand_path('~/rbbt-slurm')
+Path.setup(workdir)
+
+running_jobs = begin
+                squeue_txt = CMD.cmd('squeue').read
+                squeue_txt.split("\n").collect{|l| l.to_i.to_s}
+               rescue
+                Log.warn "Cannot determine if jobs are running, they will seem to be all alive (Job ID in green)"
+                squeue_txt = nil
+                $norunningjobs = true
+                []
+               end
+
+if squeue_txt
+ job_nodes = {}
+ squeue_txt.split("\n").each do |line|
+  parts = line.strip.split(/\s+/)
+  job_nodes[parts.first] = parts.last.split(",")
+ end
+else
+ job_nodes = nil
+end
+
+count = 0
+workdir.glob("**/command.slurm").sort_by{|f| File.mtime(f)}.each do |fcmd|
+ dir = File.dirname(fcmd)
+
+ if m = Open.read(fcmd).match(/#CMD: (.*)/)
+  cmd = m[1]
+ else
+  cmd = nil
+ end
+
+ if m = Open.read(fcmd).match(/# Run command\n(.*?)\n/im)
+  exe = m[1]
+ else
+  exe = nil
+ end
+
+ if m = Open.read(fcmd).match(/^CONTAINER_DIR=(.*)/)
+  container_home = m[1]
+ else
+  container_home = nil
+ end
+
+
+ if File.exists?(fid = File.join(dir, 'job.id'))
+  id = Open.read(fid).chomp
+ else
+  id = nil
+ end
+
+ if File.exists?(fstatus = File.join(dir, 'exit.status'))
+  exit_status = Open.read(fstatus).to_i
+ else
+  exit_status = nil
+ end
+
+ if File.exists?(fstatus = File.join(dir, 'job.status'))
+  nodes = Open.read(fstatus).split("\n").last.split(/\s+/).last.split(",")
+ elsif job_nodes[id]
+  nodes = job_nodes[id]
+ else
+  nodes = []
+ end
+
+ if File.exists?(File.join(dir, 'std.out'))
+  outt = File.mtime File.join(dir, 'std.out')
+  errt = File.mtime File.join(dir, 'std.err')
+  time_diff = Time.now - [outt, errt].max
+ end
+
+ fdep = File.join(dir, 'dependencies.list')
+ deps = Open.read(fdep).split("\n") if File.exists?(fdep)
+
+ fcadep = File.join(dir, 'canfail_dependencies.list')
+ cadeps = Open.read(fcadep).split("\n") if File.exists?(fcadep)
+
+ aborted = error = true if aborted.nil? && error.nil?
+ if done || error || aborted || running || queued || jobid || search
+   select = false
+   select = true if done && exit_status && exit_status.to_i == 0 
+   select = true if error && exit_status && exit_status.to_i != 0 
+   select = true if aborted && (exit_status.nil? && ! running_jobs.include?(id))
+   select = select && jobid.split(",").include?(id) if jobid 
+   select = select && cmd.match(/#{search}/) if search
+   next unless select 
+ end
+
+
+ puts Log.color(:yellow, "**ERASING**")
+ puts Log.color :blue, dir
+ puts Log.color(:magenta, "Creation: ") << File.mtime(File.join(dir, 'command.slurm')).to_s
+ puts Log.color(:magenta, "Done: ") << File.mtime(File.join(dir, 'exit.status')).to_s if File.exist?(File.join(dir, 'exit.status'))
+ puts Log.color(:magenta, "Exec: ") << (exe || "Missing")
+ puts Log.color(:magenta, "CMD: ") << (Log.color(:yellow, cmd) || "Missing")
+ puts Log.color(:magenta, "HOME: ") << Log.color(:yellow, container_home) if container_home
+ puts Log.color(:magenta, "Job ID: ") << (exit_status ? (exit_status == 0 ? Log.color(:green, "Done") : Log.color(:red, "Error")) + " (#{ id })" :  (running_jobs.include?(id) || $norunningjobs ? Log.color(:green, id) : Log.color(:red, id) ))
+ puts Log.color(:magenta, "Dependencies: ") << deps * ", " if deps
+ puts Log.color(:magenta, "Dependencies (can fail): ") << cadeps * ", " if cadeps
+ puts Log.color(:magenta, "Nodes: ") << nodes * ", "
+ puts Log.color(:magenta, "Output: ") << File.exists?(File.join(dir, 'std.out')).to_s << (id.nil? ? "" : " (last update " + Misc.format_seconds(time_diff) + " ago)")
+
+ if options[:sbatch_parameters]
+   puts Log.color(:magenta, "SBATCH parameters: ")
+   puts Log.color :blue, CMD.cmd('grep "^#SBATCH" |tail -n +6', :in => Open.read(fcmd)).read.strip
+ end
+
+ if tail && File.exists?(File.join(dir, 'std.err'))
+   if exit_status && exit_status != 0
+     puts Log.color(:magenta, "First error or exception found: ")
+     puts CMD.cmd("grep -i -w 'error\\|[a-z]*exception' #{File.join(dir, 'std.err')} -A #{tail.to_i} |head -n #{tail.to_i}", :no_fail => true).read
+   elsif exit_status
+     puts Log.color(:magenta, "Completed jobs: ")
+     puts CMD.cmd("grep -i -w 'Completed step' #{File.join(dir, 'std.err')} | grep -v 'Retrying dep.' | tail -n #{tail.to_i}", :no_fail => true).read
+   else
+     puts Log.color(:magenta, "Log tail: ")
+     puts CMD.cmd("tail -n #{tail.to_i} #{File.join(dir, 'std.err')}").read
+   end
+ end
+
+ count += 1
+
+ Open.rm_rf dir unless dry_run
+end
+
+puts 
+puts "Found #{count} jobs"
+

data/share/rbbt_commands/slurm/list

@@ -0,0 +1,220 @@
+#!/usr/bin/env ruby
+
+require 'rbbt-util'
+require 'rbbt/util/simpleopt'
+
+#$0 = "rbbt #{$previous_commands*""} #{ File.basename(__FILE__) }" if $previous_commands
+
+options = SOPT.setup <<EOF
+
+Queue a job in Marenostrum
+
+$ rbbt mnl [options] 
+
+-h--help Print this help
+-d--done Done jobs only
+-e--error Error jobs only
+-a--aborted SLURM aboted jobs
+-r--running Running jobs only
+-q--queued Queued jobs only
+-j--job* Job ids
+-s--search* Regular expression
+-t--tail* Show the last lines of the STDERR
+-SBP--sbatch_parameters show sbatch parameters
+-PERF--procpath_performance show Procpath performance summary
+-sacct--sacct_peformance show sacct performance summary
+EOF
+
+if options[:help]
+ if defined? rbbt_usage
+  rbbt_usage 
+ else
+  puts SOPT.doc
+ end
+ exit 0
+end
+
+Log.severity = 4
+done, error, running, queued, aborted, jobid, search, tail = options.values_at :done, :error, :running, :queued, :aborted, :job, :search, :tail
+
+workdir = File.expand_path('~/rbbt-slurm')
+Path.setup(workdir)
+
+running_jobs = begin
+                squeue_txt = CMD.cmd('squeue').read
+                squeue_txt.split("\n").collect{|l| l.to_i.to_s}
+               rescue
+                Log.warn "Cannot determine if jobs are running, they will seem to be all alive (Job ID in green)"
+                squeue_txt = nil
+                $norunningjobs = true
+                []
+               end
+
+if squeue_txt
+ job_nodes = {}
+ squeue_txt.split("\n").each do |line|
+  parts = line.strip.split(/\s+/)
+  job_nodes[parts.first] = parts.last.split(",")
+ end
+else
+ job_nodes = nil
+end
+
+count = 0
+workdir.glob("**/command.slurm").sort_by{|f| File.mtime(f)}.each do |fcmd|
+ dir = File.dirname(fcmd)
+
+ if m = Open.read(fcmd).match(/#CMD: (.*)/)
+  cmd = m[1]
+ else
+  cmd = nil
+ end
+
+ if m = Open.read(fcmd).match(/# Run command\n(.*?)\n/im)
+  exe = m[1]
+ else
+  exe = nil
+ end
+
+ if m = Open.read(fcmd).match(/^CONTAINER_DIR=(.*)/)
+  container_home = m[1]
+ else
+  container_home = nil
+ end
+
+
+ if File.exists?(fid = File.join(dir, 'job.id'))
+  id = Open.read(fid).chomp
+ else
+  id = nil
+ end
+
+ if File.exists?(fstatus = File.join(dir, 'exit.status'))
+  exit_status = Open.read(fstatus).to_i
+ else
+  exit_status = nil
+ end
+
+ if File.exists?(fstatus = File.join(dir, 'job.status'))
+  nodes = Open.read(fstatus).split("\n").last.split(/\s+/).last.split(",")
+ elsif job_nodes[id]
+  nodes = job_nodes[id]
+ else
+  nodes = []
+ end
+
+ if File.exists?(File.join(dir, 'std.out'))
+  outt = File.mtime File.join(dir, 'std.out')
+  errt = File.mtime File.join(dir, 'std.err')
+  time_diff = Time.now - [outt, errt].max
+ end
+
+ fdep = File.join(dir, 'dependencies.list')
+ deps = Open.read(fdep).split("\n") if File.exists?(fdep)
+
+ fcadep = File.join(dir, 'canfail_dependencies.list')
+ cadeps = Open.read(fcadep).split("\n") if File.exists?(fcadep)
+
+ if done || error || aborted || running || queued || jobid || search
+   select = false
+   select = true if done && exit_status == 0
+   select = true if error && exit_status && exit_status != 0 
+   select = true if aborted && (exit_status.nil? && ! running_jobs.include?(id))
+   select = true if queued && deps && (running_jobs & deps).any?
+   select = true if running && (exit_status.nil? && running_jobs.include?(id)) && (!deps || (running_jobs & deps).empty?)
+   select = true if jobid && jobid.split(",").include?(id)
+   select = true if search && cmd.match(/#{search}/)
+   next unless select 
+ end
+
+
+ puts Log.color :blue, dir
+ puts Log.color(:magenta, "Creation: ") << File.mtime(File.join(dir, 'command.slurm')).to_s
+ puts Log.color(:magenta, "Done: ") << File.mtime(File.join(dir, 'exit.status')).to_s if File.exist?(File.join(dir, 'exit.status'))
+ puts Log.color(:magenta, "Exec: ") << (exe || "Missing")
+ puts Log.color(:magenta, "CMD: ") << (Log.color(:yellow, cmd) || "Missing")
+ puts Log.color(:magenta, "HOME: ") << Log.color(:yellow, container_home) if container_home
+ puts Log.color(:magenta, "Job ID: ") << (exit_status ? (exit_status == 0 ? Log.color(:green, "Done") : Log.color(:red, "Error")) + " (#{ id })" :  (running_jobs.include?(id) || $norunningjobs ? Log.color(:green, id) : Log.color(:red, id) ))
+ puts Log.color(:magenta, "Dependencies: ") << deps * ", " if deps
+ puts Log.color(:magenta, "Dependencies (can fail): ") << cadeps * ", " if cadeps
+ puts Log.color(:magenta, "Nodes: ") << nodes * ", "
+ puts Log.color(:magenta, "Output: ") << File.exists?(File.join(dir, 'std.out')).to_s << (id.nil? ? "" : " (last update " + Misc.format_seconds(time_diff) + " ago)")
+
+ if options[:sbatch_parameters]
+   puts Log.color(:magenta, "SBATCH parameters: ")
+   text = CMD.cmd('grep "^#SBATCH" |tail -n +6', :in => Open.read(fcmd)).read.strip
+   lines = text.split("\n").collect{|line| header, _sep, value = line.partition(/\s+/); Log.color(:yellow, header + ": ") + value}
+   puts Log.color :yellow, lines * "\n"
+ end
+
+ fprocpath = File.join(dir, 'procpath.sqlite3')
+ if options[:procpath_performance]  && Open.exists?(fprocpath)
+   puts Log.color(:magenta, "Procpath summary: ")
+   require 'rbbt/tsv/csv'
+   meta = TSV.csv(CMD.cmd("sqlite3 -header -csv #{fprocpath} 'select * from meta;' "))
+   perf = TSV.csv(CMD.cmd("sqlite3 -header -csv #{fprocpath} 'select * from record;' "))
+
+   page_size = meta["page_size"].first.to_f
+   clock_ticks = meta["clock_ticks"].first.to_f
+
+   cpu_average = {}
+   rss_average = {}
+   perf.through :key, ["ts", 'stat_pid', "stat_utime", "stat_stime", "stat_cutime", "stat_cstime", "stat_rss"] do |k, values|
+    time, stat_pid, ucpu, scpu, ccpu, cscpu, rss = values
+    time = time.to_f
+
+    cpu = Misc.sum([ucpu, scpu].collect{|v| v.to_f})
+    cpu_average[stat_pid] ||= {}
+    cpu_average[stat_pid][time] ||= []
+    cpu_average[stat_pid][time] << cpu.to_f
+    rss_average[time] ||= []
+    rss_average[time] << rss.to_f * page_size
+   end
+
+   ticks = 0
+   cpu_average.each do |stat_pid, cpu_average_pid|
+    start = cpu_average_pid.keys.sort.first
+    eend = cpu_average_pid.keys.sort.last
+    ticks += Misc.sum(cpu_average_pid[eend]) - Misc.sum(cpu_average_pid[start])
+   end
+   start = rss_average.keys.sort.first
+   eend = rss_average.keys.sort.last
+   time_elapsed = eend - start
+   puts Log.color(:yellow, "CPU average: ") + "%.2f" % ( ticks / clock_ticks / time_elapsed * 100).to_s
+   puts Log.color(:yellow, "RSS average: ") + "%.2f GB" % Misc.mean(rss_average.collect{|t,l| Misc.sum(l) / (1024 * 1024 * 1024)}).to_s 
+
+ end
+
+ if options[:sacct_peformance]
+   begin
+    tsv = TSV.open(CMD.cmd("sacct -j #{id} -o 'jobid,AveRSS,MaxRSS,MaxDiskRead,MaxDiskWrite' -P|grep 'JobID\\|\.batch'"), :header_hash => '', :sep => "|", :type => :list)
+    values = tsv[tsv.keys.first]
+    if values.compact.any?
+     puts Log.color(:magenta, "SACCT performance: ")
+     puts values.zip(values.fields).collect{|v,t| Log.color(:yellow, t + ": ")  + v.to_s } * "\n"
+    end
+   rescue
+   end
+ end
+
+
+ if tail && File.exists?(File.join(dir, 'std.err'))
+   if exit_status && exit_status != 0
+     puts Log.color(:magenta, "First error or exception found: ")
+     puts CMD.cmd("grep -i -w 'error\\|[a-z]*exception' #{File.join(dir, 'std.err')} -A #{tail.to_i} |head -n #{tail.to_i}", :no_fail => true).read
+   elsif exit_status
+     puts Log.color(:magenta, "Completed jobs: ")
+     puts CMD.cmd("grep -i -w 'Completed step' #{File.join(dir, 'std.err')} | grep -v 'Retrying dep.' | tail -n #{tail.to_i}", :no_fail => true).read
+   else
+     puts Log.color(:magenta, "Log tail: ")
+     puts CMD.cmd("tail -n #{tail.to_i} #{File.join(dir, 'std.err')}").read
+   end
+ end
+
+ count += 1
+
+end
+
+puts 
+puts "Found #{count} jobs"
+