rbbt-util 5.32.7 → 5.32.13

data/share/rbbt_commands/slurm/clean

@@ -0,0 +1,212 @@
+#!/usr/bin/env ruby
+
+require 'rbbt-util'
+require 'rbbt/util/simpleopt'
+require 'rbbt/hpc'
+
+#$0 = "rbbt #{$previous_commands*""} #{ File.basename(__FILE__) }" if $previous_commands
+
+options = SOPT.setup <<EOF
+
+Clean error or aborted jobs
+
+$ rbbt slurm clean [options] 
+
+-h--help Print this help
+-d--done Done jobs only
+-e--error Error jobs only
+-a--aborted SLURM aboted jobs
+-q--queued Queued jobs only
+-j--job* Job ids
+-s--search* Regular expression
+-t--tail* Show the last lines of the STDERR
+-BP--batch_parameters show batch 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
+
+batch_system = options.delete :batch_system
+batch_system ||= 'auto'
+
+HPC::BATCH_MODULE = HPC.batch_system batch_system
+
+raise ParameterException.new("Could not detect batch_system: #{Misc.fingerprint batch_system}") if HPC::BATCH_MODULE.nil?
+
+Log.severity = 4
+done, error, aborted, queued, jobid, search, tail, batch_parameters, dry_run = options.values_at :done, :error, :aborted, :queued, :job, :search, :tail, :batch_parameters, :dry_run
+
+workdir = File.expand_path('~/rbbt-batch')
+Path.setup(workdir)
+
+running_jobs = begin
+                squeue_txt = HPC::BATCH_MODULE.job_status
+                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.batch").sort_by{|f| File.mtime(f)}.each do |fcmd|
+ dir = File.dirname(fcmd)
+ command_txt = Open.read(fcmd)
+
+ if m = command_txt.match(/#CMD: (.*)/)
+  cmd = m[1]
+ else
+  cmd = nil
+ end
+
+ if m = command_txt.match(/# Run command\n(.*?)\n/im)
+  exe = m[1]
+ else
+  exe = nil
+ end
+
+ if m = command_txt.match(/^CONTAINER_DIR=(.*)/)
+  container_home = m[1]
+ else
+  container_home = nil
+ end
+
+ if m = command_txt.match(/^BATCH_SYSTEM=(.*)/)
+  job_batch_system = m[1].downcase
+ else
+  job_batch_system = nil
+ end
+
+ different_system = job_batch_system != batch_system
+
+ 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'))
+   fstatus_txt = Open.read(fstatus) 
+   begin
+     if job_batch_system == "lsf"
+       nodes = Open.read(fstatus).split("\n").last.split(/\s+/)[5].split(",")
+     else
+       nodes = Open.read(fstatus).split("\n").last.split(/\s+/).last.split(",")
+     end
+   rescue
+     nodes = []
+   end
+ 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
+
+ if done || error || aborted || queued || jobid 
+   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))
+   is_running =  exit_status.nil? && ( (running_jobs.include?(id) && (!deps || (running_jobs & deps).empty?)) || different_system )
+   select = true if queued && deps && (running_jobs & deps).any? || queued && is_running && nodes.empty?
+   select = true if jobid && jobid.split(",").include?(id)
+   select = select && cmd.match(/#{search}/) if search
+   next unless select 
+ elsif search
+   select = false
+   select = true if search && cmd.match(/#{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.batch')).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[:batch_parameters]
+   puts Log.color(:magenta, "BATCH parameters: ")
+   case job_batch_system 
+   when 'slurm'
+     puts Log.color :blue, CMD.cmd('grep "^#SBATCH" |tail -n +6', :in => Open.read(fcmd)).read.strip
+   when 'lsf'
+     puts Log.color :blue, CMD.cmd('grep "^#BSUB" |tail -n +6', :in => Open.read(fcmd)).read.strip
+   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
+
+ Open.rm_rf dir unless dry_run
+end
+
+puts 
+puts "Found #{count} jobs"
+

data/share/rbbt_commands/slurm/list

@@ -0,0 +1,315 @@
+#!/usr/bin/env ruby
+
+require 'rbbt-util'
+require 'rbbt/util/simpleopt'
+require 'rbbt/hpc'
+
+#$0 = "rbbt #{$previous_commands*""} #{ File.basename(__FILE__) }" if $previous_commands
+
+options = SOPT.setup <<EOF
+
+Queue a job in Marenostrum
+
+$ rbbt slurm list [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
+-p--progress Report progress of job and the dependencies
+-BP--batch_parameters show batch parameters
+-BPP--batch_procpath show Procpath performance summary
+-sacct--sacct_peformance show sacct performance summary
+-bs--batch_system* Batch system to use: auto, lsf, slurm (default is auto-detect)
+EOF
+
+if options[:help]
+ if defined? rbbt_usage
+  rbbt_usage 
+ else
+  puts SOPT.doc
+ end
+ exit 0
+end
+
+batch_system = options.delete :batch_system
+batch_system ||= 'auto'
+
+HPC::BATCH_MODULE = HPC.batch_system batch_system
+
+raise ParameterException.new("Could not detect batch_system: #{Misc.fingerprint batch_system}") if HPC::BATCH_MODULE.nil?
+
+batch_system = HPC::BATCH_MODULE.to_s.split("::").last.downcase
+
+done, error, running, queued, aborted, jobid, search, tail, progress = options.values_at :done, :error, :running, :queued, :aborted, :job, :search, :tail, :progress
+
+workdir = File.expand_path('~/rbbt-batch')
+Path.setup(workdir)
+
+running_jobs = begin
+                squeue_txt = HPC::BATCH_MODULE.job_status
+                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.batch").sort_by{|f| File.mtime(f)}.each do |fcmd|
+ dir = File.dirname(fcmd)
+ command_txt = Open.read(fcmd)
+
+ if m = command_txt.match(/#CMD: (.*)/)
+  cmd = m[1]
+ else
+  cmd = nil
+ end
+
+ if m = command_txt.match(/^export BATCH_SYSTEM=(.*)/)
+  job_batch_system = m[1].downcase
+ else
+  job_batch_system = nil
+ end
+
+ different_system = job_batch_system != batch_system
+
+ if m = command_txt.match(/#MANIFEST: (.*)/)
+  manifest = m[1]
+ else
+  manifest = nil
+ end
+
+ if m = command_txt.match(/#STEP_PATH: (.*)/)
+  step_path = m[1]
+ else
+  step_path = nil
+ end
+
+ if m = command_txt.match(/#EXEC_CMD: (.*)/)
+  exe = m[1]
+ else
+  exe = nil
+ end
+
+ if m = command_txt.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'))
+   fstatus_txt = Open.read(fstatus) 
+   begin
+     if job_batch_system == "lsf"
+       nodes = Open.read(fstatus).split("\n").last.split(/\s+/)[5].split(",")
+     else
+       nodes = Open.read(fstatus).split("\n").last.split(/\s+/).last.split(",")
+     end
+   rescue
+     nodes = []
+   end
+ elsif job_nodes[id]
+   nodes = job_nodes[id].reject{|n| n.include? "("}
+ else
+   nodes = []
+ end
+
+ if File.exists?(File.join(dir, 'exit.status'))
+   now = File.ctime(File.join(dir, 'exit.status'))
+ else
+   now = Time.now
+ end
+
+ if File.exists?(File.join(dir, 'std.out'))
+  cerrt = File.ctime File.join(dir, 'std.err')
+  coutt = File.ctime File.join(dir, 'std.out')
+  outt = File.mtime File.join(dir, 'std.out')
+  errt = File.mtime File.join(dir, 'std.err')
+  time_diff = now - [outt, errt].max
+  time_elapsed = now - [cerrt, coutt].min
+ 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 
+   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))
+   is_running =  exit_status.nil? && ( (running_jobs.include?(id) && (!deps || (running_jobs & deps).empty?)) || different_system )
+   select = true if queued && deps && (running_jobs & deps).any? || queued && is_running && nodes.empty?
+   select = true if running && nodes.any? && (exit_status.nil? && running_jobs.include?(id)) && (!deps || (running_jobs & deps).empty?)
+   select = true if jobid && jobid.split(",").include?(id)
+   select = select && step_path.match(/#{search}/) if search
+   next unless select 
+ elsif search
+   select = false
+   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.batch')).to_s
+ puts Log.color(:magenta, "Started: ") << File.ctime(File.join(dir, 'std.err')).to_s if File.exist?(File.join(dir, 'std.err'))
+ puts Log.color(:magenta, "Manifest: ") << Log.color(:yellow, manifest) 
+ puts Log.color(:magenta, "Step path: ") << Log.color(:yellow, step_path)
+ 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
+ if different_system
+   puts Log.color(:magenta, "Job ID (#{Log.color(:red, job_batch_system)}): ") << (exit_status ? (exit_status == 0 ? Log.color(:green, "Done") : Log.color(:red, "Error")) + " (#{ id })" :  Log.color(:green, id)  )
+ else
+   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) ))
+ end
+ 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, "Time elapsed: ") << Misc.format_seconds(time_elapsed) if time_elapsed
+ puts Log.color(:magenta, "Output: ") << File.exists?(File.join(dir, 'std.out')).to_s << (id.nil? || File.exists?(File.join(dir, 'exit.status')) ? "" : " (last update " + Misc.format_seconds(time_diff) + " ago)")
+
+ if options[:batch_parameters]
+   puts Log.color(:magenta, "BATCH parameters: ")
+   case job_batch_system 
+   when 'slurm'
+     text = CMD.cmd('grep "^#SBATCH" ', :in => Open.read(fcmd)).read.strip
+   when 'lsf'
+     text = CMD.cmd('grep "^#BSUB" ', :in => Open.read(fcmd)).read.strip
+   else
+     text = ""
+   end
+   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[:batch_procpath]  && 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
+   ticks = 1 if ticks == 0
+   time_elapsed = 1 if time_elapsed == 0
+   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 
+   puts Log.color(:yellow, "Time: ") + Misc.format_seconds((eend - start))
+
+ end
+
+ if options[:sacct_peformance]
+   begin
+     raise "sacct not supported for LSF" unless batch_system == 'slurm' 
+     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
+     Log.warn $!.message
+   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(" cat #{File.join(dir, 'std.err')} | grep -v '^[^\\s:]*\\[3.m'  | tail -n #{tail.to_i} ").read
+   end
+ end
+
+ if options[:progress] 
+   step_line = Open.read(fcmd).split("\n").select{|line| line =~ /^#STEP_PATH:/}.first
+   if step_line
+     require 'rbbt/workflow'
+     step_path = step_line.split(": ").last.strip
+     step = Step.new step_path
+     step.load_dependencies_from_info
+     has_bar = false
+     (step.rec_dependencies + [step]).reverse.each do |j|
+       next if j.done? 
+       if j.file(:progress).exists?
+         bar = Log::ProgressBar.new 
+         bar.load(j.file(:progress).yaml)
+         puts Log.color(:magenta, "Progress: ") + bar.report_msg + " " + Log.color(:yellow, j.task_signature)
+         has_bar = true
+       end
+     end
+     puts Log.color(:magenta, "Progress: ") + Log.color(:yellow, step.task_signature) + " #{step.status}" unless has_bar
+   end
+ end
+
+ count += 1
+
+end
+
+puts 
+puts "Found #{count} jobs"
+