[feat] add compression and tests to sha1 store (#1032)

Co-authored-by: Min Xu <min.xu.public@gmail.com>

fairscale/experimental/wgit/__init__.py

@@ -3,8 +3,23 @@
 # This source code is licensed under the BSD license found in the
 # LICENSE file in the root directory of this source tree.
 
+import sys
 from typing import List
 
+# Check for user requirements before we import our code.
+try:
+    import pygit2
+except ImportError:
+    print("Error: please pip install pygit2 module to use wgit")
+    sys.exit(1)
+
+try:
+    import pgzip
+except ImportError:
+    print("Error: please pip install pgzip module to use wgit")
+    sys.exit(1)
+
+
 from .repo import Repo
 from .signal_sparsity import Algo, SignalSparsity
 from .version import __version_tuple__

fairscale/experimental/wgit/sha1_store.py

@@ -12,8 +12,9 @@ import shutil
 import sys
 import tempfile
 import time
-from typing import Any, Dict, Union, cast
+from typing import Any, Dict, Optional, Union, cast
 
+import pgzip
 import torch
 from torch import Tensor
 
@@ -25,6 +26,7 @@ SHA1_STORE_DIR_NAME = "sha1_store"
 
 # Const string keys for json file. Do not change for backward compatibilities.
 RF_KEY = "ref_count"
+COMP_KEY = "compressed"
 
 
 def _get_json_entry(d: Dict[str, Any]) -> Dict[str, Any]:
@@ -38,6 +40,28 @@ def _get_json_entry(d: Dict[str, Any]) -> Dict[str, Any]:
     return d
 
 
+def _copy_compressed(src: Path, dest: Path, thread: Optional[int], blocksize: int) -> None:
+    """Helper to copy a file and compress it at the same time."""
+    with open(str(src), "rb") as srcf:
+        with pgzip.open(str(dest), "wb", compresslevel=5, thread=thread, blocksize=blocksize) as destf:
+            while True:
+                buf = srcf.read(blocksize)
+                if len(buf) == 0:
+                    break
+                destf.write(buf)
+
+
+def _copy_uncompressed(src: Path, dest: Path, thread: Optional[int], blocksize: int) -> None:
+    """Helper to copy a file and uncompress it at the same time."""
+    with open(str(dest), "wb") as destf:
+        with pgzip.open(str(src), "rb", thread=thread, blocksize=blocksize) as srcf:
+            while True:
+                buf = srcf.read(blocksize)
+                if len(buf) == 0:
+                    break
+                destf.write(buf)
+
+
 class SHA1_Store:
     """
     This class represents a SHA1 checksum based storage dir for state_dict
@@ -61,6 +85,12 @@ class SHA1_Store:
     to delete in a version tracking graph. The lesson here is that content
     addressibility and dependency graphs do not mix well.
 
+    We support multicore compression for the data to be store on per-object basis.
+    The ``torch.save()`` API uses zip format to store the data, but it appears to
+    be uncompressed. Even if it can be made compressed, it is likely a single
+    threaded compression. Therefore, we use pgzip to do parallel
+    compression/decompression on top of it to use all the cores.
+
     Args:
         parent_path (Path):
             The parent path in which a SHA1_Store will be created.
@@ -75,16 +105,29 @@ class SHA1_Store:
         sha1_buf_size (int):
             Buffer size used for checksumming. Default: 100MB.
         tmp_dir (str):
-            Dir for temporary files if input is an in-memory object.
+            Dir for temporary files if input is an in-memory object or output data needs
+            to be decompressed first.
+        pgzip_threads (int, optional):
+            Number of threads (cores) used in compression. Default: None to use all cores.
+        pgzip_block_size (int):
+            Per-thread block size for compression. Default: 10MB.
     """
 
     def __init__(
-        self, parent_path: Path, init: bool = False, sha1_buf_size: int = 100 * 1024 * 1024, tmp_dir: str = ""
+        self,
+        parent_path: Path,
+        init: bool = False,
+        sha1_buf_size: int = 100 * 1024 * 1024,
+        tmp_dir: str = "",
+        pgzip_threads: Optional[int] = None,
+        pgzip_block_size: int = 10 * 1024 * 1024,
     ) -> None:
         """Create or wrap (if already exists) a sha1_store."""
         self._path = parent_path.joinpath(SHA1_STORE_DIR_NAME)
         self._ref_file_path = self._path.joinpath("ref_count.json")
         self._sha1_buf_size = sha1_buf_size
+        self._pgzip_threads = pgzip_threads
+        self._pgzip_block_size = pgzip_block_size
         self._json_dict: Dict[str, Any] = {"created_on": time.ctime()}
 
         # Initialize the sha1_store if not exist and init==True.
@@ -121,7 +164,7 @@ class SHA1_Store:
         with open(self._ref_file_path, "w", encoding="utf-8") as f:
             json.dump(self._json_dict, f, ensure_ascii=False, indent=4)
 
-    def add(self, file_or_obj: Union[Path, Tensor, OrderedDict]) -> str:
+    def add(self, file_or_obj: Union[Path, Tensor, OrderedDict], compress: bool = False) -> str:
         """
         Adds a file/object to the internal sha1_store and the sha1 references
         accordingly.
@@ -130,6 +173,9 @@ class SHA1_Store:
         in <file_or_obj> is moved within the sha1_store and the reference file is updated.
         If the input is an object, it will be store in the self._tmp_dir and then moved.
 
+        If compress is True, the stored file is also compressed, which is useful for tensors
+        with a lot of zeros.
+
         Args:
             file_or_obj (str or tensor or OrderedDict):
                 Path to the file to be added to the sha1_store or an in-memory object
@@ -155,7 +201,7 @@ class SHA1_Store:
         sha1_hash = self._get_sha1_hash(file_path)
 
         # Add reference.
-        ref_count = self._add_ref(sha1_hash, True)
+        ref_count = self._add_ref(sha1_hash, True, compress)
 
         if ref_count == 1:
             # First time adding
@@ -172,12 +218,15 @@ class SHA1_Store:
             # Transfer the file to the internal sha1_store
             repo_fpath = repo_fdir.joinpath(sha1_hash)
             try:
-                shutil.copy2(file_path, repo_fpath)
+                if compress:
+                    _copy_compressed(file_path, repo_fpath, self._pgzip_threads, self._pgzip_block_size)
+                else:
+                    shutil.copy2(file_path, repo_fpath)
             except BaseException as error:
                 # Something went wrong, perhaps out of space, or race condition due to lack of locking.
                 # TODO (Min): proper handle the error and recover when we learn more here.
                 sys.stderr.write(f"An exception occured: {repr(error)}\n")
-                ref_count = self._add_ref(sha1_hash, False)
+                ref_count = self._add_ref(sha1_hash, False, compress)
 
         # Clean up if needed.
         if remove_tmp:
@@ -210,7 +259,18 @@ class SHA1_Store:
         #
         # TODO (Min): we could also keep a stats in the meta data on how many
         #             times the object is read. Will add if that's needed.
-        return torch.load(path)
+        self._load_json_dict()
+        if self._json_dict[sha1][COMP_KEY]:
+            # Compressed. Because pgzip doesn't support tell() yet, we need to
+            # uncompress into a temp file and return it.
+            tmp = self._get_tmp_file_path()
+            _copy_uncompressed(path, tmp, self._pgzip_threads, self._pgzip_block_size)
+            obj = torch.load(tmp)
+            tmp.unlink()
+            return obj
+        else:
+            # Uncompressed.
+            return torch.load(path)
 
     def delete(self, sha1: str) -> None:
         """Delete a SHA1
@@ -282,7 +342,7 @@ class SHA1_Store:
         part1, part2 = sha1[:2], sha1[2:4]
         return self._path.joinpath(part1, part2)
 
-    def _add_ref(self, current_sha1_hash: str, inc: bool) -> int:
+    def _add_ref(self, current_sha1_hash: str, inc: bool, compressed: bool) -> int:
         """
         Update the reference count.
 
@@ -312,6 +372,9 @@ class SHA1_Store:
         entry[RF_KEY] += 1 if inc else -1
         assert entry[RF_KEY] >= 0, "negative ref count"
 
+        # Update compressed flag.
+        entry[COMP_KEY] = compressed
+
         self._json_dict[current_sha1_hash] = entry
         self._store_json_dict()
 

requirements-dev.txt

@@ -34,5 +34,8 @@ numpy == 1.22.0
 # For layerwise gradient scaler
 sklearn >= 0.0
 
-# For weigit
+# For weigit. These are actually user requirements, not developer requirements.
+# However, due to the experimental nature of weigit, we don't expose to the
+# general users of fairscale yet. We check for them in weigit's init code.
 pygit2==1.9.2
+pgzip==0.3.1

tests/experimental/wgit/test_sha1_store.py

@@ -47,7 +47,8 @@ def sha1_store(request):
     return sha1_store
 
 
-def test_sha1_add_file(sha1_store):
+@pytest.mark.parametrize("compress", [True, False])
+def test_sha1_add_file(sha1_store, compress):
     os.chdir(PARENT_DIR)
 
     # Create random checkpoints
@@ -65,15 +66,15 @@ def test_sha1_add_file(sha1_store):
 
     # Add those 5 random files.
     for c in chkpts:
-        sha1_store.add(c)
+        sha1_store.add(c, compress)
 
     # Add a fixed data twice.
     module = nn.Linear(100, 100, bias=False)
     module.weight.data = torch.zeros(100, 100)
     zeros_file = "zeros.pt"
     torch.save(module.state_dict(), zeros_file)
-    sha1_store.add(zeros_file)
-    sha1_store.add(zeros_file)
+    sha1_store.add(zeros_file, compress)
+    sha1_store.add(zeros_file, compress)
 
     # Assert the ref counts are 1,1,1,1,1 and 2
     sha1_store._load_json_dict()
@@ -86,16 +87,17 @@ def test_sha1_add_file(sha1_store):
     assert sorted(map(lambda x: x["ref_count"], json_dict.values())) == [1, 1, 1, 1, 1, 2], json_dict
 
 
-def test_sha1_add_state_dict(sha1_store):
+@pytest.mark.parametrize("compress", [True, False])
+def test_sha1_add_state_dict(sha1_store, compress):
     os.chdir(PARENT_DIR)
     # add once
     for i in range(3):
-        sha1_store.add(nn.Linear(10, 10).state_dict())
+        sha1_store.add(nn.Linear(10, 10).state_dict(), compress)
     # add twice
     for i in range(3):
         sd = nn.Linear(8, 8).state_dict()
-        sha1_store.add(sd)
-        sha1_store.add(sd)
+        sha1_store.add(sd, compress)
+        sha1_store.add(sd, compress)
 
     sha1_store._load_json_dict()
     json_dict = sha1_store._json_dict
@@ -103,9 +105,10 @@ def test_sha1_add_state_dict(sha1_store):
     assert sorted(map(lambda x: x["ref_count"], json_dict.values())) == [1, 1, 1, 2, 2, 2], json_dict
 
 
-def test_sha1_add_tensor(sha1_store):
+@pytest.mark.parametrize("compress", [True, False])
+def test_sha1_add_tensor(sha1_store, compress):
     os.chdir(PARENT_DIR)
-    sha1_store.add(torch.Tensor([1.0, 5.5, 3.4]))
+    sha1_store.add(torch.Tensor([1.0, 5.5, 3.4]), compress)
     sha1_store._load_json_dict()
     json_dict = sha1_store._json_dict
     if torch_version() >= (1, 9, 0):
@@ -114,7 +117,8 @@ def test_sha1_add_tensor(sha1_store):
         assert key in json_dict.keys() and json_dict[key]["ref_count"] == 1, json_dict
 
 
-def test_sha1_get(sha1_store):
+@pytest.mark.parametrize("compress", [True, False])
+def test_sha1_get(sha1_store, compress):
     """Testing the get() API: normal and exception cases."""
     os.chdir(PARENT_DIR)
 
@@ -125,15 +129,15 @@ def test_sha1_get(sha1_store):
     tensor = torch.ones(20, 30)
 
     # Check that we can get them back.
-    file_sha1 = sha1_store.add(file)
+    file_sha1 = sha1_store.add(file, compress)
     sd = sha1_store.get(file_sha1)
     assert objects_are_equal(sd, torch.load(file))
 
-    sd_sha1 = sha1_store.add(state_dict)
+    sd_sha1 = sha1_store.add(state_dict, compress)
     sd = sha1_store.get(sd_sha1)
     assert objects_are_equal(sd, state_dict)
 
-    tensor_sha1 = sha1_store.add(tensor)
+    tensor_sha1 = sha1_store.add(tensor, compress)
     tensor_got = sha1_store.get(tensor_sha1)
     assert objects_are_equal(tensor_got, tensor)
 
@@ -142,22 +146,23 @@ def test_sha1_get(sha1_store):
         sha1_store.get(tensor_sha1[:-1])
 
 
-def test_sha1_delete(sha1_store):
+@pytest.mark.parametrize("compress", [True, False])
+def test_sha1_delete(sha1_store, compress):
     """Testing the delete() API: with ref counting behavior."""
     os.chdir(PARENT_DIR)
 
     # Add once and delete, second delete should throw an exception.
     tensor = torch.ones(30, 50)
-    sha1 = sha1_store.add(tensor)
+    sha1 = sha1_store.add(tensor, compress)
     sha1_store.delete(sha1)
     with pytest.raises(ValueError):
         sha1_store.delete(sha1)
 
     # Add multiple times and delete should match that.
     state_dict = nn.Sequential(nn.Linear(10, 10), nn.Linear(10, 20)).state_dict()
-    sha1 = sha1_store.add(state_dict)
+    sha1 = sha1_store.add(state_dict, compress)
     for i in range(3):
-        new_sha1 = sha1_store.add(state_dict)
+        new_sha1 = sha1_store.add(state_dict, compress)
         assert sha1 == new_sha1, f"{sha1} vs. {new_sha1}"
     for i in range(4):
         sha1_store.delete(sha1)