1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
|
#!/usr/bin/env python
# SPDX-FileCopyrightText: Copyright (c) 2021-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: LicenseRef-NvidiaProprietary
#
# NVIDIA CORPORATION, its affiliates and licensors retain all intellectual
# property and proprietary rights in and to this material, related
# documentation and any modifications thereto. Any use, reproduction,
# disclosure or distribution of this material and related documentation
# without an express license agreement from NVIDIA CORPORATION or
# its affiliates is strictly prohibited.
import gpustats
import time
class GPULowUtilization(gpustats.GPUOperation):
DEFAULT_THRESHOLD = 50
DEFAULT_NUM_CHUNKS = 30
display_name = 'DEPRECATED - Use gpu_time_util instead'
usage = '{SCRIPT} -- {{DISPLAY_NAME}}'
should_display = False
message_advice = ("The following are time regions with an average GPU"
" utilization below {THRESHOLD}%%. Addressing the gaps might improve"
" application performance.\n\n"
"Suggestions:\n"
" 1. Use CPU sampling data, OS Runtime blocked state backtraces,"
" and/or OS Runtime APIs related to thread synchronization to"
" understand if a sluggish or blocked CPU is causing the gaps.\n"
" 2. Add NVTX annotations to CPU code to understand the reason"
" behind the gaps.")
message_noresult = ("There were no problems detected with GPU utilization."
" No time regions have an average GPU utilization below {THRESHOLD}%%.")
def MessageAdvice(self, extended=True):
return self.message_advice.format(
THRESHOLD=self._threshold, NUM_CHUNKS=self._chunks)
def MessageNoResult(self):
return self.message_noresult.format(
THRESHOLD=self._threshold, NUM_CHUNKS=self._chunks)
create_chunk_table = """
CREATE TEMP TABLE CHUNK (
rangeId INTEGER PRIMARY KEY NOT NULL
)
"""
insert_chunk_table = """
INSERT INTO temp.CHUNK
WITH RECURSIVE
range AS (
SELECT
0 AS rangeId
UNION ALL
SELECT
rangeId + 1 AS rangeId
FROM
range
LIMIT {NUM_CHUNKS}
)
SELECT rangeId FROM range
"""
query_format_columns = """
SELECT
ROW_NUMBER() OVER(ORDER BY average, duration) AS "Row#",
average AS "In-Use:ratio_%",
duration AS "Duration:dur_ns",
start AS "Start:ts_ns",
pid AS "PID",
deviceId AS "Device ID",
contextId AS "Context ID",
globalId AS "_Global ID",
api AS "_API"
FROM
({GPU_UNION_TABLE})
LIMIT {ROW_LIMIT}
"""
# Return chunks that have an average GPU utilization below the given threshold.
# 1. CTE "range": Define the range being analyzed for each deviceId/PID with
# the corresponding chunk size.
# 2. CTE "chunk": Duplicate chunks for each deviceId/PID with the appropriate
# start and end.
# 3. CTE "utilization": Find all ranges being run in each chunk and keep only
# the ones that have a percentage of GPU utilization lower than the threshold.
# If there are multiple streams, the utilizations are added up.
# 4. CTE "chunkgroup": Give a groupId that will be used to define consecutive
# chunks.
# 5. Coalesce chunks with same groupId and calculate the weighted average.
query_chunk = """
WITH
ops AS (
{{GPU_TABLE}}
),
range AS (
SELECT
min(start) AS start,
max(end) AS end,
ceil(CAST(max(end) - min(start) AS FLOAT) / {NUM_CHUNKS}) AS chunkSize,
pid,
globalId,
deviceId,
contextId,
api
FROM
ops
GROUP BY deviceId, pid
),
chunk AS (
SELECT
chunk.rangeId,
chunk.rangeId * range.chunkSize + range.start AS cstart,
min(chunk.rangeId * range.chunkSize + range.start + range.chunkSize, range.end) AS cend,
chunkSize,
range.pid,
range.globalId,
range.deviceId,
range.contextId,
range.api
FROM
temp.CHUNK AS chunk
JOIN
range
ON cstart < cend
),
utilization AS (
SELECT
chunk.rangeId,
chunk.cstart AS start,
chunk.cend AS end,
chunk.cend - chunk.cstart AS size,
sum(CAST(coalesce(min(ops.end, chunk.cend) - max(ops.start, chunk.cstart), 0) AS FLOAT)) / (chunk.cend - chunk.cstart) * 100 AS timePercentage,
chunk.pid,
chunk.globalId,
chunk.deviceId,
chunk.contextId,
chunk.api
FROM
chunk
LEFT JOIN
ops
ON ops.deviceId == chunk.deviceId
AND ops.pid == chunk.pid
AND ops.start < chunk.cend
AND ops.end > chunk.cstart
GROUP BY
chunk.rangeId, chunk.deviceId, chunk.pid
HAVING
timePercentage < {THRESHOLD}
),
chunkgroup AS
(
SELECT
*,
rangeId - ROW_NUMBER() OVER (PARTITION BY pid, deviceId ORDER BY rangeId) AS groupId
FROM
utilization
)
SELECT
min(start) AS start,
max(end) - min(start) AS duration,
round(sum(size * timePercentage) / sum(size), 1) AS average,
pid,
globalId,
deviceId,
contextId,
api
FROM
chunkgroup
GROUP BY groupId, deviceId, pid
LIMIT {ROW_LIMIT}
"""
_arg_opts = [
[['threshold'], {'type': int, 'default': DEFAULT_THRESHOLD,
'help': 'maximum percentage of time the GPU is being used'}],
[['chunks'], {'type': int, 'default': DEFAULT_NUM_CHUNKS,
'help': 'number of equal-duration chunks'}],
]
def setup(self):
err = super().setup()
if err != None:
return err
self._threshold = self.parsed_args.threshold
self._chunks = self.parsed_args.chunks
if self._chunks and not 1 <= self._chunks <= 1000:
self._parser.error("argument --chunks: value must be between 1 and 1000")
self.statements = [
self.create_chunk_table,
self.insert_chunk_table.format(NUM_CHUNKS = self._chunks)]
err = self.create_gpu_ops_view(self.query_chunk.format(
NUM_CHUNKS = self._chunks,
THRESHOLD = self._threshold,
ROW_LIMIT = self._row_limit))
if err != None:
return err
self.query = self.query_format_columns.format(
GPU_UNION_TABLE = self.query_gpu_ops_union(),
ROW_LIMIT = self._row_limit)
if __name__ == "__main__":
GPULowUtilization.Main()
|
