Re: [LTP] [Patch 2/6] Developed new functions and fixed issues causing ilb test failure

[poornima nayak <mpnayak@linux.vnet.ibm.com>]

On Wed, 2009-10-14 at 14:55 +0530, Gautham R Shenoy wrote:
> On Tue, Oct 13, 2009 at 01:13:46PM +0530, Poornima Nayak wrote:
> > +def is_multi_core():
> > +    ''' Return true if system has sockets has multiple cores
> > +    '''
> > +  
> > +    try:
> > +        file_cpuinfo = open("/proc/cpuinfo", 'r')
> > +        for line in file_cpuinfo:
> > +            if line.startswith('siblings'):
> > +                siblings = line.split(":")
> > +            if line.startswith('cpu cores'):
> > +                cpu_cores = line.split(":")
> > +                break
> 
> I assume your code works only on x86 machine for now.
This part of the code is modified to read the information from sys
topology. For this months LTP release patches will be mailed.
> 
> The corresponding kernel code found from arch/x86/kernel/cpu/proc.c is
> as follows
> ========================================================================
> /*
>  *	Get CPU information for use by the procfs.
>  */
> static void show_cpuinfo_core(struct seq_file *m, struct cpuinfo_x86 *c,
> 			      unsigned int cpu)
> {
> #ifdef CONFIG_SMP
> 	if (c->x86_max_cores * smp_num_siblings > 1) {
> 		seq_printf(m, "physical id\t: %d\n", c->phys_proc_id);
> 		seq_printf(m, "siblings\t: %d\n",
> 			   cpumask_weight(cpu_core_mask(cpu)));
> 		seq_printf(m, "core id\t\t: %d\n", c->cpu_core_id);
> 		seq_printf(m, "cpu cores\t: %d\n", c->booted_cores);
> 		seq_printf(m, "apicid\t\t: %d\n", c->apicid);
> 		seq_printf(m, "initial apicid\t: %d\n", c->initial_apicid);
> 	}
> #endif
> }
> ============================================================================
> IIUC, 'siblings' tells us the number of hardware threads available
> within a socket and 'cpu cores' tells us the number of cores available
> within a socket.
> 
> So, if you are checking if the socket has multiple cores or not, 'cpu
> cores > 1' would suffice, no ? Why do we need the following code below ?
> 
This code is also modified to read information from sys topology.
Testcase will lear wether system is hyper threaded by reading bit mask
of cpus in thread_siblings file.
> > +       
> > +        if int( siblings[1] ) == int( cpu_cores[1] ):
> 		^^^^^^^^^^^^^^^^^^^^^
> 		This only means that each core in the socket has only
> 		one processing unit or thread.
> > +            if int( cpu_cores[1] ) > 1:
> > +                multi_core = 1
> > +            else:
> > +                multi_core = 0
> > +        else:
> > +            num_of_cpus = int(siblings[1]) / int(cpu_cores[1])
> 		^^^^^^^^^^
> 		This tells us the number of hardware threads per core.
> > +            if num_of_cpus > 1:
> > +                multi_core = 1
> > +            else:
> > +                multi_core = 0
> > +        file_cpuinfo.close()
> > +        return multi_core
> > +    except Exception:
> > +        print "Failed to check if system is multi core system"
> > +        sys.exit(1)
> > +
> >  def get_hyper_thread_count():
> >      ''' Return number of threads in CPU. For eg for x3950 this function
> >          would return 2. In future if 4 threads are supported in CPU, this
> > @@ -153,6 +184,40 @@ def map_cpuid_pkgid():
> >                  sys.exit(1)
> > 
> > 
> > +def generate_sibling_list():
> > +    ''' Routine to generate siblings list
> > +    '''
> > +    try:
> > +        for i in range(0, cpu_count):
> > +            siblings_file = '/sys/devices/system/cpu/cpu%s' % i
> > +            siblings_file += '/topology/thread_siblings_list'
> > +            threads_sibs = open(siblings_file).read().rstrip()
> > +            thread_ids = threads_sibs.split("-")
> A string '0-3' means that 0,1,2,3 are the siblings and not just 0 and 3.
> Also, you can have strings like '0,4' which means that only 0 and 4 are
> siblings.
> 
Modified this function. Entire function is modified to read cpu bit mask
information from thread_siblings. 
> Thus I am not sure if the parsing that you're doing is sufficient
> to generate the list.
> 
> > +    
> > +            if not thread_ids in siblings_list:
> > +                siblings_list.append(thread_ids)
> > +    except Exception, details:
> > +        print "Exception in generate_siblings_list", details
> > +        sys.exit(1)
> > +
> > +def get_siblings(cpu_id):
> > +    ''' Return siblings of cpu_id
> > +    '''
> > +    try:
> > +        cpus = ""
> > +        for i in range(0, len(siblings_list)):
> > +            for cpu in siblings_list[i]:
> > +                if cpu_id == cpu:
> > +                    for j in siblings_list[i]:
> > +                        # Exclude cpu_id in the list of siblings
> > +                        if j != cpu_id:
> > +                            cpus += j
> > +                    return cpus
> > +        return cpus
> > +    except Exception, details:
> > +        print "Exception in get_siblings", details
> > +        sys.exit(1)
> > +
> >  def get_proc_data(stats_list):
> >      ''' Read /proc/stat info and store in dictionary
> >      '''
> > @@ -168,18 +233,18 @@ def get_proc_data(stats_list):
> >          sys.exit(1)
> > 
> >  def get_proc_loc_count(loc_stats):
> > -    ''' Read /proc/stat info and store in dictionary
> > +    ''' Read /proc/interrupts info and store in list
> >      '''
> >      try:
> >          file_procstat = open("/proc/interrupts", 'r')
> >          for line in file_procstat:
> > -            if line.startswith('LOC:'):
> > +            if line.startswith(' LOC:') or line.startswith('LOC:'):
> >                  data = line.split()
> >                  for i in range(0, cpu_count):
> >                      # To skip LOC
> >                      loc_stats.append(data[i+1])
> > -                    print data[i+1]
> > -        file_procstat.close()
> > +                file_procstat.close()
> > +                return
> >      except Exception, details:
> >          print "Could not read interrupt statistics", details
> >          sys.exit(1)
> > @@ -192,6 +257,8 @@ def set_sched_mc_power(sched_mc_level):
> >          os.system('echo %s > \
> >              /sys/devices/system/cpu/sched_mc_power_savings 2>/dev/null'
> >              % sched_mc_level)
> > +
> > +        get_proc_data(stats_start)
> >      except OSError, e:
> >          print "Could not set sched_mc_power_savings to", sched_mc_level, e
> >  	sys.exit(1)
> > @@ -203,6 +270,8 @@ def set_sched_smt_power(sched_smt_level)
> >          os.system('echo %s > \
> >              /sys/devices/system/cpu/sched_smt_power_savings 2>/dev/null'
> >              % sched_smt_level)
> > +
> > +        get_proc_data(stats_start)
> >      except OSError, e:
> >          print "Could not set sched_smt_power_savings to", sched_smt_level, e
> >  	sys.exit(1)
> > @@ -218,21 +287,36 @@ def set_timer_migration_interface(value)
> >          print "Could not set timer_migration to ", value, e
> >          sys.exit(1)
> > 
> > -def trigger_ebizzy (sched_smt, stress, duration, background, pinned):
> > -    ''' Triggers ebizzy workload for sched_mc=1
> > -        testing
> > +def get_job_count(stress, workload, sched_smt):
> > +    ''' Returns number of jobs/threads to be triggered
> >      '''
> > +    
> >      try:
> >          if stress == "thread":
> >              threads = get_hyper_thread_count()
> 
> I am assuming get_hyper_thread_count() returns the number of
> hyperthreads within a core.
Yes number of hyper threads within a core
> 
> So, if stress = "thread", the number of software threads we create is
> equal to the numbef of h/w threads within a core ?
yes number of software threads is equal to number of hyper threads
within a core. In a system like mx3950 where every core has two threads,
workload will be triggered with two threads.
> 
> 
> >          if stress == "partial":
> >              threads = cpu_count / socket_count
> 
> If stress = "partial", the number of software threads we create is equal
> to the number of h/w threads within a socket ?
if stress is partial, then software threads we create is equal number of
cores / 2. 
> 
> > +            if is_hyper_threaded():
> > +                if workload == "ebizzy" and int(sched_smt) ==0:
> > +                    threads = threads / get_hyper_thread_count()
> 
> Thus if the workload is ebizzy and sched_smt is 0, the number of
> software threads we create is equal to the number of cores within the
> socket, right ?
Yes number of cores within a socket. But is system is hyper threaded
then number of software threads is further divided by hyper thread
count. Else jobs will be triggered in other socket.
> 
> > +                if workload == "kernbench" and int(sched_smt) < 2:
> > +                    threads = threads / get_hyper_thread_count()    
> >          if stress == "full":
> > -	    threads = cpu_count
> > +            threads = cpu_count
> >          if stress == "single_job":
> >              threads = 1
> >              duration = 180
> > +        return threads
> > +    except Exception, details:
> > +        print "get job count failed ", details
> > +        sys.exit(1)
> > 
> > +def trigger_ebizzy (sched_smt, stress, duration, background, pinned):
> > +    ''' Triggers ebizzy workload for sched_mc=1
> > +        testing
> > +    '''
> > +    try:
> > +        threads = get_job_count(stress, "ebizzy", sched_smt)
> >          olddir = os.getcwd()
> >          path = '%s/utils/benchmark' % os.environ['LTPROOT']
> >          os.chdir(path)
> > @@ -282,23 +366,14 @@ def trigger_ebizzy (sched_smt, stress, d
> >          print "Ebizzy workload trigger failed ", details
> >          sys.exit(1)   
> > 
> > -def trigger_kernbench (sched_smt, stress, background, pinned):
> > +def trigger_kernbench (sched_smt, stress, background, pinned, perf_test):
> >      ''' Trigger load on system like kernbench.
> >          Copys existing copy of LTP into as LTP2 and then builds it
> >          with make -j
> >      '''
> >      olddir = os.getcwd()
> >      try:
> > -        if stress == "thread":
> > -	    threads = 2
> > -        if stress == "partial":
> > -	    threads = cpu_count / socket_count
> > -            if is_hyper_threaded() and int(sched_smt) !=2:
> > -                threads = threads / get_hyper_thread_count()
> > -        if stress == "full":
> > -            threads = cpu_count
> > -        if stress == "single_job":
> > -            threads = 1
> > +        threads = get_job_count(stress, "kernbench", sched_smt)
> > 
> >          dst_path = "/root"
> >          olddir = os.getcwd()      
> > @@ -335,24 +410,35 @@ def trigger_kernbench (sched_smt, stress
> >          get_proc_loc_count(intr_start)
> >          if pinned == "yes":
> >              os.system ( 'taskset -c %s %s/kernbench -o %s -M -H -n 1 \
> > -                >/dev/null 2>&1' % (cpu_count-1, benchmark_path, threads))
> > +                >/dev/null 2>&1 &' % (cpu_count-1, benchmark_path, threads))
> > +
> > +            # We have to delete import in future
> > +            import time
> > +            time.sleep(240)
> > +            stop_wkld("kernbench")
> >          else:
> >              if background == "yes":
> >                  os.system ( '%s/kernbench -o %s -M -H -n 1 >/dev/null 2>&1 &' \
> >                      % (benchmark_path, threads))
> >              else:
> > -                os.system ( '%s/kernbench -o %s -M -H -n 1 >/dev/null 2>&1' \
> > -                    % (benchmark_path, threads))
> > +                if perf_test == "yes":
> > +                    os.system ( '%s/kernbench -o %s -M -H -n 1 >/dev/null 2>&1' \
> > +                        % (benchmark_path, threads))
> > +                else:
> > +                    os.system ( '%s/kernbench -o %s -M -H -n 1 >/dev/null 2>&1 &' \
> > +                        % (benchmark_path, threads))
> > +                    # We have to delete import in future
> > +                    import time
> > +                    time.sleep(240)
> > +                    stop_wkld("kernbench")
> >          
> >          print "INFO: Workload kernbench triggerd"
> >          os.chdir(olddir)
> > -        #get_proc_data(stats_stop)
> > -        #get_proc_loc_count(intr_stop)
> >      except Exception, details:
> >          print "Workload kernbench trigger failed ", details
> >          sys.exit(1)
> >     
> > -def trigger_workld(sched_smt, workload, stress, duration, background, pinned):
> > +def trigger_workld(sched_smt, workload, stress, duration, background, pinned, perf_test):
> >      ''' Triggers workload passed as argument. Number of threads 
> >          triggered is based on stress value.
> >      '''
> > @@ -360,7 +446,7 @@ def trigger_workld(sched_smt, workload, 
> >          if workload == "ebizzy":
> >              trigger_ebizzy (sched_smt, stress, duration, background, pinned)
> >          if workload == "kernbench":
> > -            trigger_kernbench (sched_smt, stress, background, pinned)
> > +            trigger_kernbench (sched_smt, stress, background, pinned, perf_test)
> >      except Exception, details:
> >          print "INFO: Trigger workload failed", details
> >          sys.exit(1)
> > @@ -434,7 +520,7 @@ def generate_report():
> >              print >> keyvalfile, "package-%s=%3.4f" % \
> >  		(pkg, (float(total_idle)*100/total))
> >      except Exception, details:
> > -        print "Generating reportfile failed: ", details
> > +        print "Generating utilization report failed: ", details
> >          sys.exit(1)
> > 
> >      #Add record delimiter '\n' before closing these files
> > @@ -454,20 +540,18 @@ def generate_loc_intr_report():
> > 
> >          get_proc_loc_count(intr_stop)
> > 
> > -        print "Before substracting"
> > -        for i in range(0, cpu_count):
> > -            print "CPU",i, intr_start[i], intr_stop[i]
> > -            reportfile = open('/procstat/cpu-loc_interrupts', 'a')
> > -            print >> reportfile, "=============================================="
> > -            print >> reportfile, "     Local timer interrupt stats              "
> > -            print >> reportfile, "=============================================="
> > +        reportfile = open('/procstat/cpu-loc_interrupts', 'a')
> > +        print >> reportfile, "=============================================="
> > +        print >> reportfile, "     Local timer interrupt stats              "
> > +        print >> reportfile, "=============================================="
> > +
> >          for i in range(0, cpu_count):
> >              intr_stop[i] =  int(intr_stop[i]) - int(intr_start[i])
> >              print >> reportfile, "CPU%s: %s" %(i, intr_stop[i])
> >          print >> reportfile
> >          reportfile.close()
> >      except Exception, details:
> > -        print "Generating reportfile failed: ", details
> > +        print "Generating interrupt report failed: ", details
> >          sys.exit(1)
> > 
> >  def record_loc_intr_count():
> > @@ -542,25 +626,24 @@ def validate_cpugrp_map(cpu_group, sched
> >                                  modi_cpu_grp.remove(core_cpus[i]) 
> >                                  if len(modi_cpu_grp) == 0:
> >                                      return 0
> > -                            else:
> > +                            #This code has to be deleted 
> > +                            #else:
> >                                  # If sched_smt == 0 then its oky if threads run
> >                                  # in different cores of same package 
> > -                                if sched_smt_level == 1:
> > -                                    sys.exit(1)
> > -                                else:
> > -                                    if len(cpu_group) == 2 and \
> > -                                        len(modi_cpu_grp) < len(cpu_group):
> > -                                        print "INFO:CPUs utilized not in a core"
> > -                                        return 1                                        
> > -            print "INFO: CPUs utilized is not in same package or core"
> > -            return(1)
> > +                                #if sched_smt_level > 0 :
> > +                                    #return 1
> >  	else:
> >              for pkg in sorted(cpu_map.keys()):
> >                  pkg_cpus = cpu_map[pkg]
> > -                if pkg_cpus == cpu_group:
> > -                    return(0)
> > -                 
> > -            return(1) 
> > +                if len(cpu_group) == len(pkg_cpus):
> > +                    if pkg_cpus == cpu_group:
> > +                        return(0)
> > +                else:
> > +                    if int(cpus_utilized[0]) in cpu_map[pkg] or int(cpus_utilized[1]) in cpu_map[pkg]:
> > +                        return(0)
> > +
> > +        return(1) 
> > +
> >      except Exception, details:
> >          print "Exception in validate_cpugrp_map: ", details
> >          sys.exit(1)
> > @@ -605,36 +688,70 @@ def verify_sched_domain_dmesg(sched_mc_l
> >          print "Reading dmesg failed", details
> >          sys.exit(1)
> > 
> > -def validate_cpu_consolidation(work_ld, sched_mc_level, sched_smt_level):
> > +def get_cpu_utilization(cpu):
> > +    ''' Return cpu utilization of cpu_id
> > +    '''
> > +    try:
> > +        for l in sorted(stats_percentage.keys()):
> 
> What's the key used to index the elements of this hash table ?
> 
> 
> > +            if cpu == stats_percentage[l][0]:
> > +                return stats_percentage[l][1]
> > +        return -1
> > +    except Exception, details:
> > +        print "Exception in get_cpu_utilization", details
> > +        sys.exit(1)
> > +
> > +def validate_cpu_consolidation(stress, work_ld, sched_mc_level, sched_smt_level):
> >      ''' Verify if cpu's on which threads executed belong to same
> >      package
> >      '''
> >      cpus_utilized = list()
> > +    threads = get_job_count(stress, work_ld, sched_smt_level)
> >      try:
> >          for l in sorted(stats_percentage.keys()):
> >              #modify threshold
> > +            cpu_id = stats_percentage[l][0].split("cpu")
> > +            if cpu_id[1] == '':
> > +                continue
> > +            if int(cpu_id[1]) in cpus_utilized:
> > +                continue
> >              if is_hyper_threaded():
> Does is_hyper_threaded() check the status of the system from procfs or
> sysfs every time it's called ? An easier way could be to obtain that
> once and save the state in a variable, if it's not done that way
> already.
It is read from sysfs hereafter. will look into the code and modify.
> 
> > -                if stats_percentage[l][1] > 25 and work_ld == "kernbench":
> > -                    cpu_id = stats_percentage[l][0].split("cpu")
> > -                    if cpu_id[1] != '':
> > +                if work_ld == "kernbench" and sched_smt_level < sched_mc_level:
> > +                    siblings = get_siblings(cpu_id[1])
> > +                    if siblings != "":
> > +                        sib_list = siblings.split()
> > +                        utilization = int(stats_percentage[l][1])
> > +                        for i in range(0, len(sib_list)):
> > +                            utilization += int(get_cpu_utilization("cpu%s" %sib_list[i])) 
> > +                    else:
> > +                        utilization = stats_percentage[l][1]
> > +                    if utilization > 40:
> >                          cpus_utilized.append(int(cpu_id[1]))
> > +                        if siblings != "":
> > +                            for i in range(0, len(sib_list)):
> > +                                cpus_utilized.append(int(sib_list[i]))
> >                  else:
> > -                    if stats_percentage[l][1] > 70:
> > -                        cpu_id = stats_percentage[l][0].split("cpu")
> > -                        if cpu_id[1] != '':
> > -                            cpus_utilized.append(int(cpu_id[1]))
> > +                    # This threshold wuld be modified based on results
> 
> It's easier to have these constants stored in some variables with
> appropriate names. That would be much easier to edit and update later
> instead of searching the code every time.
This modification is done.
> 
> > +                    if stats_percentage[l][1] > 40:
> > +                        cpus_utilized.append(int(cpu_id[1]))
> >              else:
> > -                if stats_percentage[l][1] > 70:
> > -                    cpu_id = stats_percentage[l][0].split("cpu")
> > -                    if cpu_id[1] != '':
> > +                if work_ld == "kernbench" :
> > +                    if stats_percentage[l][1] > 50:
> >                          cpus_utilized.append(int(cpu_id[1]))
> > -                    cpus_utilized.sort()
> > +                else:
> > +                    if stats_percentage[l][1] > 70:
> > +                        cpus_utilized.append(int(cpu_id[1]))
> > +            cpus_utilized.sort()
> >          print "INFO: CPU's utilized ", cpus_utilized
> > 
> > +        # If length of CPU's utilized is not = number of jobs exit with 1
> > +        if len(cpus_utilized) < threads:
> > +            return 1
> > +
> >          status = validate_cpugrp_map(cpus_utilized, sched_mc_level, \
> >              sched_smt_level)
> >          if status == 1:
> >              print "INFO: CPUs utilized is not in same package or core"
> > +
> >          return(status)
> >      except Exception, details:
> >          print "Exception in validate_cpu_consolidation: ", details
> > @@ -645,7 +762,8 @@ def get_cpuid_max_intr_count():
> >      try:
> >          highest = 0
> >          second_highest = 0
> > -        global cpu1_max_intr, cpu2_max_intr
> > +        cpus_utilized = []
> > +        
> >          #Skipping CPU0 as it is generally high
> >          for i in range(1, cpu_count):
> >              if int(intr_stop[i]) > int(highest):
> > @@ -658,15 +776,19 @@ def get_cpuid_max_intr_count():
> >                  if int(intr_stop[i]) > int(second_highest):
> >                      second_highest = int(intr_stop[i])
> >                      cpu2_max_intr = i
> > +        cpus_utilized.append(cpu1_max_intr)
> > +        cpus_utilized.append(cpu2_max_intr)
> > +        
> >          for i in range(1, cpu_count):
> >              if i != cpu1_max_intr and i != cpu2_max_intr:
> >                  diff = second_highest - intr_stop[i]
> >                  ''' Threshold of difference has to be manipulated '''
> >                  if diff < 10000:
> >                      print "INFO: Diff in interrupt count is below threshold"
> > -                    return 1
> > +                    cpus_utilized = []
> > +                    return cpus_utilized
> >          print "INFO: Interrupt count in other CPU's low as expected"
> > -        return 0 
> > +        return cpus_utilized
> >      except Exception, details:
> >          print "Exception in get_cpuid_max_intr_count: ", details
> >          sys.exit(1)
> > @@ -675,14 +797,12 @@ def validate_ilb (sched_mc_level, sched_
> >      ''' Validate if ilb is running in same package where work load is running
> >      '''
> >      try:
> > -        status = get_cpuid_max_intr_count()
> > -        if status == 1:
> > +        cpus_utilized = get_cpuid_max_intr_count()
> > +        if not cpus_utilized:
> >              return 1
> > -        for pkg in sorted(cpu_map.keys()):
> > -            if cpu1_max_intr in cpu_map[pkg] and cpu2_max_intr in cpu_map[pkg]:
> > -                return 0
> > -        print "INFO: CPUs with higher interrupt count is not in same package"
> > -        return 1
> > +       
> > +        status = validate_cpugrp_map(cpus_utilized, sched_mc_level, sched_smt_level)
> > +        return status
> >      except Exception, details:
> >          print "Exception in validate_ilb: ", details
> >          sys.exit(1)
> > @@ -706,3 +826,14 @@ def reset_schedsmt():
> >      except OSError, e:
> >          print "Could not set sched_smt_power_savings to 0", e
> >          sys.exit(1)
> > +
> > +def stop_wkld(work_ld):
> > +    ''' Kill workload triggered in background
> > +    '''
> > +    try:
> > +        os.system('pkill %s 2>/dev/null' %work_ld)
> > +        if work_ld == "kernbench":
> > +            os.system('pkill make 2>/dev/null')
> > +    except OSError, e:
> > +        print "Exception in stop_wkld", e
> > +        sys.exit(1)
> 


------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day 
trial. Simplify your report design, integration and deployment - and focus on 
what you do best, core application coding. Discover what's new with
Crystal Reports now.  http://p.sf.net/sfu/bobj-july
_______________________________________________
Ltp-list mailing list
Ltp-list@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/ltp-list