Merge pull request #411 from ldbc/emr-tweaks

Add option to control number of nodes & partitions

Author: szarnyasg

tools/emr/submit_datagen_job.py

@@ -9,23 +9,23 @@
 import __main__
 
 from math import ceil
-from botocore.credentials import subprocess
 from datagen import lib, util
-import subprocess
 
 import argparse
 
 from datagen.util import KeyValue, split_passthrough_args
 
 min_num_workers = 1
 max_num_workers = 1000
+min_num_threads = 1
 
 defaults = {
     'bucket': 'ldbc-snb-datagen-store',
     'use_spot': True,
     'master_instance_type': 'r6gd.2xlarge',
-    'instance_type': 'r6gd.4xlarge',
-    'sf_ratio': 100.0,  # ratio of SFs and machines. a ratio of 250.0 for SF1000 yields 4 machines
+    'instance_type': 'i3.4xlarge',
+    'sf_per_executor': 3e3,
+    'sf_per_partition': 10,
     'az': 'us-west-2c',
     'yes': False,
     'ec2_key': None,
@@ -43,15 +43,6 @@
     ec2_instances = [dict(row) for row in reader]
 
 
-def calculate_cluster_config(scale_factor, sf_ratio, vcpu):
-    num_workers = max(min_num_workers, min(max_num_workers, ceil(scale_factor / sf_ratio)))
-    num_threads = ceil(num_workers * vcpu * 2)
-    return {
-        'num_workers': num_workers,
-        'num_threads': num_threads
-    }
-
-
 def get_instance_info(instance_type):
     def parse_vcpu(col):
         return int(re.search(r'(\d+) .*', col).group(1))
@@ -76,7 +67,10 @@ def submit_datagen_job(name,
                        jar,
                        use_spot,
                        instance_type,
-                       sf_ratio,
+                       executors,
+                       sf_per_executor,
+                       partitions,
+                       sf_per_partition,
                        master_instance_type,
                        az,
                        emr_release,
@@ -97,10 +91,6 @@ def submit_datagen_job(name,
     else:
         copy_filter = f'.*{build_dir}/{copy_filter}'
 
-    exec_info = get_instance_info(instance_type)
-
-    cluster_config = calculate_cluster_config(sf, sf_ratio, exec_info['vcpu'])
-
     emr = boto3.client('emr')
 
     ts = datetime.utcnow()
@@ -115,8 +105,15 @@ def submit_datagen_job(name,
         'maximizeResourceAllocation': 'true'
     }
 
+    if executors is None:
+        executors = max(min_num_workers, min(max_num_workers, ceil(sf / sf_per_executor)))
+
+    if partitions is None:
+        partitions = max(min_num_threads, ceil(sf / sf_per_partition))
+
     spark_defaults_config = {
         'spark.serializer': 'org.apache.spark.serializer.KryoSerializer',
+        'spark.default.parallelism': str(partitions),
         **(dict(conf) if conf else {})
     }
 
@@ -157,7 +154,7 @@ def submit_datagen_job(name,
                     'Market': market,
                     'InstanceRole': 'CORE',
                     'InstanceType': instance_type,
-                    'InstanceCount': cluster_config['num_workers'],
+                    'InstanceCount': executors,
                 }
             ],
             **ec2_key_dict,
@@ -178,7 +175,7 @@ def submit_datagen_job(name,
                     'Args': ['spark-submit', '--class', lib.main_class, jar_url,
                              '--output-dir', build_dir,
                              '--scale-factor', str(sf),
-                             '--num-threads', str(cluster_config['num_threads']),
+                             '--num-threads', str(partitions),
                              '--mode', mode,
                              '--format', format,
                              *passthrough_args
@@ -263,6 +260,26 @@ def submit_datagen_job(name,
                             nargs='+',
                             action=KeyValue,
                             help="SparkConf as key=value pairs")
+    executor_args=parser.add_mutually_exclusive_group()
+    executor_args.add_argument("--executors",
+                               type=int,
+                               help=f"Total number of Spark executors."
+                               )
+    executor_args.add_argument("--sf-per-executor",
+                               type=float,
+                               default=defaults['sf_per_executor'],
+                               help=f"Number of scale factors per Spark executor. Default: {defaults['sf_per_executor']}"
+                               )
+    partitioning_args = parser.add_mutually_exclusive_group()
+    partitioning_args.add_argument("--partitions",
+                                   type=int,
+                                   help=f"Total number of Spark partitions to use when generating the dataset."
+                                   )
+    partitioning_args.add_argument("--sf-per-partition",
+                                   type=float,
+                                   default=defaults['sf_per_partition'],
+                                   help=f"Number of scale factors per Spark partitions. Default: {defaults['sf_per_partition']}"
+                                   )
 
     parser.add_argument('--', nargs='*', help='Arguments passed to LDBC SNB Datagen', dest="arg")
 
@@ -271,6 +288,5 @@ def submit_datagen_job(name,
     args = parser.parse_args(self_args)
 
     submit_datagen_job(passthrough_args=passthrough_args,
-                       sf_ratio=defaults['sf_ratio'],
                        master_instance_type=defaults['master_instance_type'],
                        **args.__dict__)