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
|
# Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions
# are met:
# * Redistributions of source code must retain the above copyright
# notice, this list of conditions and the following disclaimer.
# * Redistributions in binary form must reproduce the above copyright
# notice, this list of conditions and the following disclaimer in the
# documentation and/or other materials provided with the distribution.
# * Neither the name of NVIDIA CORPORATION nor the names of its
# contributors may be used to endorse or promote products derived
# from this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
# PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
# OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
import NvRules
from RequestedMetrics import MetricRequest, RequestedMetricsParser, Importance
requested_metrics_base = [
MetricRequest("device__attribute_compute_capability_major", "cc_major"),
MetricRequest("device__attribute_compute_capability_minor", "cc_minor"),
]
requested_metrics = [
MetricRequest("lts__t_sectors_srcunit_tex.avg.pct_of_peak_sustained_elapsed"),
MetricRequest("lts__t_sectors_srcunit_tex_lookup_miss.sum"),
MetricRequest("lts__t_sectors_srcunit_tex_aperture_peer_lookup_miss.sum"),
MetricRequest("lts__t_sectors_srcunit_tex_aperture_sysmem_lookup_miss.sum"),
# additional metrics for speedup estimation
MetricRequest("dram__bytes.sum.per_second", "dram_bandwidth", Importance.OPTIONAL, 0),
MetricRequest("pcie__read_bytes.sum.per_second", "pcie_read_bandwidth", Importance.OPTIONAL, 0),
MetricRequest("pcie__write_bytes.sum.per_second", "pcie_write_bandwidth", Importance.OPTIONAL, 0),
MetricRequest("nvlrx__bytes.sum.per_second", "nvlink_read_bandwidth", Importance.OPTIONAL, 0, False),
MetricRequest("nvltx__bytes.sum.per_second", "nvlink_write_bandwidth", Importance.OPTIONAL, 0, False),
]
def get_identifier():
return "MemoryApertureUsage"
def get_name():
return "Memory Aperture Usage"
def get_description():
return "Detection of frequent memory accesses backed by apertures with slower memory bandwidth and higher latency."
def get_section_identifier():
return "MemoryWorkloadAnalysis_Chart"
def get_parent_rules_identifiers():
return ["Memory"]
def get_estimated_speedup(metrics, aperture):
all_lookup_misses = metrics["lts__t_sectors_srcunit_tex_lookup_miss.sum"].value()
aperture_lookup_misses = metrics["lts__t_sectors_srcunit_tex_aperture_{}_lookup_miss.sum".format(aperture)].value()
dram_bandwidth = metrics["dram_bandwidth"].value()
pcie_bandwidth = metrics["pcie_read_bandwidth"].value() + metrics["pcie_write_bandwidth"].value()
nvlink_bandwidth = metrics["nvlink_read_bandwidth"].value() + metrics["nvlink_write_bandwidth"].value()
if aperture == "sysmem":
# System memory is expected to be connected via PCIe
aperture_bandwidth = pcie_bandwidth
elif aperture == "peer":
# Peer memory is expected to be connected via PCIe or NVLink
aperture_bandwidth = max(pcie_bandwidth, nvlink_bandwidth)
else:
# unknown aperture, cannot calculate speedup
return NvRules.IFrontend.SpeedupType_LOCAL, 0
if all_lookup_misses != 0 and dram_bandwidth != 0 and aperture_bandwidth != 0:
# Only give an estimate if we could collect some value for the aperture bandwidth
improvement_percent = (aperture_lookup_misses / all_lookup_misses) * (1 - aperture_bandwidth / dram_bandwidth) * 100
speedup_type = NvRules.IFrontend.SpeedupType_GLOBAL
else:
improvement_percent = 0
speedup_type = NvRules.IFrontend.SpeedupType_LOCAL
return speedup_type, improvement_percent
def apply(handle):
ctx = NvRules.get_context(handle)
action = ctx.range_by_idx(0).action_by_idx(0)
fe = ctx.frontend()
metrics_base = RequestedMetricsParser(handle, action).parse(requested_metrics_base)
cc = metrics_base["cc_major"].value() * 10 + metrics_base["cc_minor"].value()
if (False
or cc == 72
or cc == 87
):
return
apertures = {
"peer" : (
"Peer"
),
"sysmem" : (
"System"
)
}
metrics = RequestedMetricsParser(handle, action).parse(requested_metrics)
lts__t_sectors_srcunit_tex_peak_pct = metrics["lts__t_sectors_srcunit_tex.avg.pct_of_peak_sustained_elapsed"].value()
lts__t_sectors_srcunit_tex_lookup_miss = metrics["lts__t_sectors_srcunit_tex_lookup_miss.sum"].value()
lts__high_utilization_threshold = 50
lts__high_aperture_utilization_threshold = 40
for aperture in apertures:
aperture_info = apertures[aperture]
metric_name = "lts__t_sectors_srcunit_tex_aperture_{}_lookup_miss.sum".format(aperture)
lts__t_sectors_srcunit_tex_aperture_lookup_miss = metrics[metric_name].value()
lts__t_sectors_srcunit_tex_aperture_lookup_miss_ratio = 100. * lts__t_sectors_srcunit_tex_aperture_lookup_miss / lts__t_sectors_srcunit_tex_lookup_miss if lts__t_sectors_srcunit_tex_lookup_miss else 0.
if lts__t_sectors_srcunit_tex_peak_pct > lts__high_utilization_threshold and lts__t_sectors_srcunit_tex_aperture_lookup_miss_ratio > lts__high_aperture_utilization_threshold:
message = "{} memory backs {:.1f}% of the data in the L2 cache that was requested by L1TEX and had cache misses in L2. ".format(aperture_info, lts__t_sectors_srcunit_tex_aperture_lookup_miss_ratio)
message += "Fetching data from {} memory is considerably slower than accessing the device's dedicated DRAM, as the data needs to be communicated over PCIE or NVLINK. ".format(aperture_info.lower())
message += "Consider moving frequently accessed data to DRAM before launching this kernel."
if 80 <= cc:
message += " Tweaking the L2 cache policies can help optimizing the cache hit rates for accesses to slower {} memory. ".format(aperture_info.lower())
message += "Lookup CUaccessProperty and policy CU_ACCESS_PROPERTY_PERSISTING for more details."
msg_id = fe.message(NvRules.IFrontend.MsgType_MSG_OPTIMIZATION, message, "{} Memory Usage".format(aperture_info))
speedup_type, speedup_value = get_estimated_speedup(metrics, aperture)
fe.speedup(msg_id, speedup_type, speedup_value)
fe.focus_metric(msg_id, metric_name, lts__t_sectors_srcunit_tex_aperture_lookup_miss, NvRules.IFrontend.Severity_SEVERITY_DEFAULT, "Decrease the lookup misses to {} memory".format(aperture_info.lower()))
|
