add pointer to script used for generating final runbook

benchmark/datasets.py

@@ -193,6 +193,13 @@ def get_dataset_iterator(self, bs=512, split=(1,0)):
             j1 = min(j0 + bs, i1)
             yield sanitize(x[j0:j1])
 
+    def get_data_in_range(self, start, end):
+        assert start >= 0
+        assert end <= self.nb
+        filename = self.get_dataset_fn()
+        x = xbin_mmap(filename, dtype=self.dtype, maxn=self.nb)
+        return x[start:end]
+
     def search_type(self):
         return "knn"
 
@@ -434,6 +441,28 @@ def distance(self):
 
     def prepare(self, skip_data=False, original_size=10 ** 9):
         return super().prepare(skip_data, original_size = self.nb)
+
+class MSTuringClustered30M(DatasetCompetitionFormat):
+    def __init__(self):
+        self.nb = 29998994
+        self.d = 100
+        self.nq = 10000
+        self.dtype = "float32"
+        self.ds_fn = "30M-clustered64.fbin"
+        self.qs_fn = "testQuery10K.fbin"
+        self.gt_fn = "clu_msturing30M_gt100"
+
+        self.base_url = "https://comp21storage.blob.core.windows.net/publiccontainer/comp23/clustered_data/msturing-30M-clustered/"
+        self.basedir = os.path.join(BASEDIR, "MSTuring-30M-clustered")
+
+        self.private_gt_url = None
+        self.private_qs_url = None
+
+    def distance(self):
+        return "euclidean"
+
+    def prepare(self, skip_data=False, original_size=10 ** 9):
+        return super().prepare(skip_data, original_size = self.nb)
 
 class MSSPACEV1B(DatasetCompetitionFormat):
     def __init__(self, nb_M=1000):
@@ -984,6 +1013,7 @@ def __str__(self):
     'msturing-1M': lambda : MSTuringANNS(1),
 
     'msturing-10M-clustered': lambda: MSTuringClustered10M(),
+    'msturing-30M-clustered': lambda: MSTuringClustered30M(),
 
     'msspacev-1B': lambda : MSSPACEV1B(1000),
     'msspacev-100M': lambda : MSSPACEV1B(100),

benchmark/results.py

@@ -6,7 +6,6 @@
 import re
 import traceback
 
-
 def get_result_filename(dataset=None, count=None, definition=None,
                         query_arguments=None, neurips23track=None, runbook_path=None):
     d = ['results']
@@ -41,9 +40,7 @@ def get_result_filename(dataset=None, count=None, definition=None,
 
 def add_results_to_h5py(f, search_type, results, count, suffix = ''):
     if search_type == "knn" or search_type == "knn_filtered":
-        neighbors = f.create_dataset('neighbors' + suffix, (len(results), count), 'i')
-        for i, idxs in enumerate(results):
-            neighbors[i] = idxs
+        neighbors = f.create_dataset('neighbors' + suffix, (len(results), count), 'i', data = results)
     elif search_type == "range":
         lims, D, I= results
         f.create_dataset('neighbors' + suffix, data=I)
@@ -59,7 +56,7 @@ def store_results(dataset, count, definition, query_arguments,
     head, tail = os.path.split(fn)
     if not os.path.isdir(head):
         os.makedirs(head)
-    f = h5py.File(fn, 'w')
+    f = h5py.File(name=fn, mode='w', libver='latest')
     for k, v in attrs.items():
         f.attrs[k] = v
 
@@ -83,7 +80,7 @@ def load_all_results(dataset=None, count=None, neurips23track=None, runbook_path
             if os.path.splitext(fn)[-1] != '.hdf5':
                 continue
             try:
-                f = h5py.File(os.path.join(root, fn), 'r+')
+                f = h5py.File(name=os.path.join(root, fn), mode='r+', libver='latest')
                 properties = dict(f.attrs)
                 yield properties, f
                 f.close()

benchmark/runner.py

@@ -100,7 +100,7 @@ def run(definition, dataset, count, run_count, rebuild,
                     algo.set_query_arguments(*query_arguments)
                 if neurips23track == 'streaming':
                     descriptor, results = custom_runner.run_task(
-                        algo, ds, distance, 1, run_count, search_type, private_query, runbook)
+                        algo, ds, distance, count, 1, search_type, private_query, runbook)
                 else:
                     descriptor, results = custom_runner.run_task(
                         algo, ds, distance, count, run_count, search_type, private_query)
@@ -116,9 +116,11 @@ def run(definition, dataset, count, run_count, rebuild,
                         X = ds.get_private_queries()
                     power_stats = power_capture.run(algo, X, distance, count,
                                                     run_count, search_type, descriptor)
+                print('start store results')
                 store_results(dataset, count, definition,
                               query_arguments, descriptor,
                               results, search_type, neurips23track, runbook_path)
+                print('end store results')
     finally:
         algo.done()
 
@@ -263,7 +265,7 @@ def run_docker(definition, dataset, count, runs, timeout, rebuild,
 
     client = docker.from_env()
     if mem_limit is None:
-        mem_limit = psutil.virtual_memory().available
+        mem_limit = psutil.virtual_memory().available if neurips23track != 'streaming' else (8*1024*1024*1024)
 
     # ready the container object invoked later in this function
     container = None

benchmark/streaming/compute_gt.py

@@ -26,7 +26,7 @@ def gt_dir(ds, runbook_path):
     return os.path.join(ds.basedir, str(ds.nb), runbook_filename)
 
 def output_gt(ds, ids, step, gt_cmdline, runbook_path):
-    data = ds.get_dataset()
+    data = ds.get_data_in_range(0, ds.nb)
     data_slice = data[ids]
 
     dir = gt_dir(ds, runbook_path)
@@ -52,6 +52,9 @@ def output_gt(ds, ids, step, gt_cmdline, runbook_path):
     gt_cmdline += ' --tags_file ' + tags_file
     print("Executing cmdline: ", gt_cmdline)
     os.system(gt_cmdline)
+    print("Removing data file")
+    rm_cmdline = "rm " + data_file
+    os.system(rm_cmdline)
 
 
 def main():

data_export.py

@@ -95,7 +95,10 @@ def cleaned_run_metric(run_metrics):
             print(f"Looking at track:{track}, dataset:{dataset_name}")
             dataset = DATASETS[dataset_name]()
             if track == 'streaming':
-                for runbook_path in ['neurips23/streaming/simple_runbook.yaml', 'neurips23/streaming/clustered_runbook.yaml', 'neurips23/streaming/delete_runbook.yaml']:
+                for runbook_path in ['neurips23/streaming/simple_runbook.yaml', 
+                                    'neurips23/streaming/clustered_runbook.yaml',
+                                    'neurips23/streaming/delete_runbook.yaml',
+                                    'neurips23/streaming/final_runbook.yaml']:
                     results = load_all_results(dataset_name, neurips23track=track, runbook_path=runbook_path)
                     run_metrics = compute_metrics_all_runs(dataset, dataset_name, results, args.recompute, \
                         args.sensors, args.search_times, args.private_query, \

neurips23/README.md

@@ -23,7 +23,7 @@ The Practical Vector Search challenge at NeurIPS 2023 has four different tasks:
     The tags are from a vocabulary of 200386 possible tags. 
     The 100,000 queries consist of one image embedding and one or two tags that must appear in the database elements to be considered.
 
-**Task Streaming:** This task uses 10M slice of the MS Turing data set released in the previous challenge. The index starts with zero points and must implement the "runbook" provided - a sequence of insertion operations, deletion operations, and search commands (roughly 4:4:1 ratio) - within a time bound of 1 hour. In the final run, we will  use a different runbook, and possibly a different data set, to avoid participants over-fitting to this dataset. Entries will be ranked by average recall over queries at all check points. The intention is for the algorithm to  process the operations and maintain a compact index over the active points rather than index the entire anticipated set of points and use tombstones or flags to mark active elements.
+**Task Streaming:** This task uses 10M slice of the MS Turing data set released in the previous challenge. The index starts with zero points and must implement the "runbook" provided - a sequence of insertion operations, deletion operations, and search commands (roughly 4:4:1 ratio) - within a time bound of 1 hour and a DRAM limit of 8GB. Entries will be ranked by average recall over queries at all check points. The intention is for the algorithm to  process the operations and maintain a compact index over the active points rather than index the entire anticipated set of points and use tombstones or flags to mark active elements. ~~In the final run, we will  use a different runbook, and possibly a different data set, to avoid participants over-fitting to this dataset.~~ The final run will use `msturing-30M-clustered`, a 30M slice of the MSTuring dataset, and the `final_runbook.yaml` runbook generated with the `final_rubook_gen.py` script. 
 
 **Task Out-Of-Distribution:**  Yandex Text-to-Image 10M represents a cross-modal dataset where the database and query vectors have different distributions in the shared vector space.
     The base set is a 10M subset of the Yandex visual search database of 200-dimensional image embeddings which are produced with the Se-ResNext-101 model. 
@@ -46,6 +46,7 @@ The baselines were run on an Azure Standard D8lds v5 (8 vcpus, 16 GiB memory) ma
 |Sparse   | Linear Scan | 101                         |  `python3 run.py --dataset sparse-full --algorithm linscan --neurips23track sparse` |
 |Filter   | faiss       | 3200                        | `python3 run.py --dataset yfcc-10M --algorithm faiss --neurips23track filter` |
 |Streaming| DiskANN     | 0.924 (recall@10), 23 mins  |  `python3 run.py --dataset msturing-10M-clustered --algorithm diskann --neurips23track streaming --runbook_path neurips23/streaming/delete_runbook.yaml` |
+|Streaming| DiskANN     | 0.883 (recall@10), 45 mins  |  `python3 run.py --dataset msturing-30M-clustered --algorithm diskann --neurips23track streaming --runbook_path neurips23/streaming/final_runbook.yaml` |
 |OOD      | DiskANN     | 4882                        | `python3 run.py --dataset text2image-10M --algorithm diskann --neurips23track ood` | 
 
 
@@ -110,13 +111,17 @@ For the competition dataset, run commands mentioned in the table above, for exam
 python run.py --neurips23track filter    --algorithm faiss   --dataset yfcc-10M
 python run.py --neurips23track sparse    --algorithm linscan --dataset sparse-full
 python run.py --neurips23track ood       --algorithm diskann --dataset text2image-10M
+# preliminary runbook for testing 
 python run.py --neurips23track streaming --algorithm diskann --dataset msturing-10M-clustered --runbook_path neurips23/streaming/delete_runbook.yaml
+#Final runbook for evaluation
+python run.py --neurips23track streaming --algorithm diskann --dataset msturing-30M-clustered --runbook_path neurips23/streaming/final_runbook.yaml
 ```
 
 For streaming track, runbook specifies the order of operations to be executed by the algorithms. To download the ground truth for every search operation: (needs azcopy tool in your binary path):
 ```
-python benchmark/streaming/download_gt.py --runbook_file neurips23/streaming/simple_runbook.yaml --dataset msspacev-10M 
-python benchmark/streaming/download_gt.py --runbook_file neurips23/streaming/delete_runbook.yaml --dataset msturing-10M-clustered 
+python -m benchmark.streaming.download_gt --runbook_file neurips23/streaming/simple_runbook.yaml --dataset msspacev-10M
+python -m benchmark.streaming.download_gt --runbook_file neurips23/streaming/delete_runbook.yaml --dataset msturing-10M-clustered
+python -m benchmark.streaming.download_gt --runbook_file neurips23/streaming/final_runbook.yaml  --dataset msturing-30M-clustered
 ```
 Alternately, to compute ground truth for an arbitrary runbook, [clone and build DiskANN repo](https://github.com/Microsoft/DiskANN) and use the command line tool to compute ground truth at various search checkpoints. The `--gt_cmdline_tool` points to the directory with DiskANN commandline tools.
 ```

neurips23/streaming/diskann/config.yaml

@@ -83,4 +83,18 @@ msturing-10M-clustered:
           args: |
             [{"R":64, "L":50, "insert_threads":16, "consolidate_threads":16}]
           query-args: |
-            [{"Ls":100, "T":16}]
+            [{"Ls":100, "T":16}]
+msturing-30M-clustered:
+    diskann: 
+      docker-tag: neurips23-streaming-diskann
+      module: neurips23.streaming.diskann.diskann-str
+      constructor: diskann
+      base-args: ["@metric"]
+      run-groups:
+        base:
+          args: |
+            [{"R":32, "L":50, "insert_threads":16, "consolidate_threads":16},
+             {"R":32, "L":70, "insert_threads":16, "consolidate_threads":16},
+             {"R":50, "L":50, "insert_threads":16, "consolidate_threads":16}]
+          query-args: |
+            [{"Ls":70, "T":16}]