Update code

Author: Aetf

CMakeLists.txt

@@ -148,6 +148,10 @@ if(WITH_TIMEOUT_WARNING)
     set(SALUS_ENABLE_TIMEOUT_WARNING 1)
 endif(WITH_TIMEOUT_WARNING)
 
+if(USE_TENSORFLOW)
+    set(SALUS_ENABLE_TENSORFLOW 1)
+endif(USE_TENSORFLOW)
+
 configure_file(src/config.h.in ${CMAKE_CURRENT_BINARY_DIR}/config.h)
 include_directories(${CMAKE_CURRENT_BINARY_DIR})
 

benchmarks/driver/runner.py

@@ -213,7 +213,8 @@ def __call__(self, executor, output_file):
         else:
             output_file.parent.mkdir(exist_ok=True, parents=True)
             with output_file.open('w') as f:
-                return execute(cmd, cwd=str(cwd), env=env, stdout=f, stderr=sp.STDOUT)
+                # return execute(cmd, cwd=str(cwd), env=env, stdout=f, stderr=sp.STDOUT)
+                return execute(cmd, cwd=str(cwd), env=env, stdout=f, stderr=None)
 
     def _construct_test_name(self, executor):
         # type: (Executor) -> Tuple[str, str]
@@ -239,7 +240,7 @@ def _construct_test_name(self, executor):
             })
         }
 
-        variable_batch_size_models = {'vae', 'superres'}
+        variable_batch_size_models = {'vae', 'superres', 'seq2seq', 'mnistsf', 'mnistcv', 'mnistlg'}
         if remove_suffix(self.wl.name, 'eval') not in variable_batch_size_models:
             if self.wl.batch_size not in self.wl.wtl.available_batch_sizes():
                 raise ValueError(f"Batch size `{self.wl.batch_size}' is not supported for {self.wl.name},"
@@ -273,6 +274,8 @@ def _construct_test_name(self, executor):
             }
 
         postfix = names.get(self.wl.batch_size, '0')
+        if model_name == 'seq2seq' and postfix == '0':
+            postfix = '2_large'
 
         method = f'{cls}.{prefix}{postfix}'
         return pkg, method

benchmarks/driver/server/config/__init__.py

@@ -21,7 +21,7 @@
 from builtins import super
 
 from absl import flags
-from copy import copy
+from copy import deepcopy
 
 from ...utils import maybe_path
 from ...utils.compatiblity import pathlib
@@ -77,7 +77,7 @@ def __setattr__(self, key, value):
     def copy(self, **kwargs):
         # type: (...) -> SalusConfig
         """Return a new copy of the tuple"""
-        return copy(self).update(**kwargs)
+        return deepcopy(self).update(**kwargs)
 
     def update(self, d=None, **kwargs):
         # type: (...) -> SalusConfig

benchmarks/exps/__init__.py

@@ -23,6 +23,8 @@
 import time
 import re
 import logging
+import string
+import random
 from absl import flags
 from typing import Union, Iterable, List, TypeVar, Callable, Optional
 
@@ -379,3 +381,8 @@ def release_on_pipe(pipe):
 def sync_on_pipe(pipe):
     wait_on_pipe(pipe)
     release_on_pipe(pipe)
+
+
+def random_id(size=6, chars=string.ascii_uppercase + string.digits):
+    """Generate a random ID"""
+    return ''.join(random.choice(chars) for _ in range(size))

benchmarks/exps/smtracing.py

@@ -170,6 +170,31 @@ def tfmps(argv):
                    )
 
 
+def train_alone(argv):
+    """Run training workload alone take note of SM usage"""
+    sm_factors = [float(v) for v in argv]
+    if not sm_factors:
+        sm_factors = [1.0, 1.5, 2.0, 2.5, 3.0]
+
+    logger.info(f"Running Salus with sm factors: {sm_factors}")
+
+    # run salus
+    for factor in sm_factors:
+        with tempfile.TemporaryDirectory() as td:
+            scfg = maybe_forced_preset(presets.OpTracing)
+            scfg.logconf = 'smtracing'
+            scfg.extra_args += [
+                '--sm-factor', f'{factor:.2f}'
+            ]
+            logger.info(f"Running Salus with sm factor: {factor}")
+            # the background training job
+            wl, pipe = create_train(Executor.Salus, 0, td)
+            run_seq(scfg.copy(output_dir=FLAGS.save_dir / "alone" / f"{factor:.2f}"),
+                    wl,
+                    RunFn(lambda *args, **kwargs: wait_on_pipe(pipe)),
+                    RunFn(lambda *args, **kwargs: release_on_pipe(pipe)))
+
+
 @case_switch_main
 def main():
-    return salus, tfmps
+    return salus, tfmps, train_alone, salus_factor