Use the Metrics definition files provided by the VTune Amplifier to create your own event-based metrics and define analyses describing which metrics should be viewed together and how they should be arranged. Because the events to be collected must be configured prior to collection, if a group of metrics is to be viewed together in an VTune Amplifier viewpoint, all events referenced by all of those metrics must be collected in the analysis type corresponding to that viewpoint. Thus, an analysis definition is used to produce both a viewpoint and an analysis type, meant for use together.
To create a custom event-based sampling analysis with user-defined metrics:
Prerequisites: Extract the content for <install_dir>\sdk\user_metrics\user_metrics.zip to a writable directory or share on your system.
The default installation path for the VTune Amplifier XE is [Program Files]\Intel\VTune Amplifier XE 2013. For the VTune Amplifier for Systems, the installation path is [Program Files]\Intel\System Studio <version>.<package_num>\VTune Amplifier <version> for Systems.
In the directory where you located the content from the
user_metrics.ziparchive, choose a Python script file corresponding to your architecture/processor type (for example,corei7_metrics.pyfor 32nm Intel® Core™ i7 processor. )Add any new metrics, use, or remove the existing ones.
Define an analysis type.
Save your changes.
Example of a Metric Definition
In the VTune Amplifier Metric definition files (for example, corei7_metrics.py), each metric is added by creating an instance of the metric class and filling in various fields on that instance. For example, the following snippet defines and fills in a metric with the ID DIVActive:
DIVActive = metric("DIV Active")
DIVActive.formula["corei7"] = ( event("ARITH.FPU_DIV_ACTIVE") / query("Clockticks"))
DIVActive.issue_eval["corei7"] = ( formula() > .05 ) * ( query("PMUHotspot") * .05 )where the first line defines a metric name, the second - a metric type and processor it is used with. The last line contains the metric formula. Default metrics in the file may give you a hint.
Example of an Analysis Definition
The following snippet defines the General Exploration analysis for Intel® microarchitecture code name Sandy Bridge:
SNBGeneralExploration = analysis("SNB General Exploration")
SNBGeneralExploration.name = "General Exploration"
SNBGeneralExploration.long_name = "Intel(R) Microarchitecture Code Name Sandy Bridge - General Exploration"
SNBGeneralExploration.description = ""
SNBGeneralExploration.valid_architectures = ["snb"]
SNBGeneralExploration.alias_name = "snb_general-exploration"
SNBGeneralExploration.always_collect["snb"] = [ event("CPU_CLK_UNHALTED.THREAD"), event("CPU_CLK_UNHALTED.REF_TSC"), event("INST_RETIRED.ANY") ]
SNBGeneralExploration.always_collect["snb"] = [
event("OFFCORE_RESPONSE.ANY_REQUEST.LLC_MISS_LOCAL.DRAM_0"),
event("OFFCORE_RESPONSE.ANY_REQUEST.LLC_MISS_LOCAL.DRAM_1") ]
SNBGeneralExploration.metric_tree = [
( query("CPI"), set(["summary","grid"]),
[ ( query("Derived Query Clockticks"), set(["summary","grid","srcasm"]) ),
( query("Instructions Retired"), set(["summary","grid","srcasm"]) ) ] ),
( "Filled Pipeline Slots", set(["summary","grid"]),
[ ( query("Retired Pipeline Slots"), set(["summary","grid","srcasm","collapsed"]),
[ ( query("Assists"), set(["summary","grid","srcasm"]) ) ] ),
( query("Cancelled Pipeline Slots"), set(["summary","grid","srcasm","collapsed"]),
[ ( query("Branch Mispredict"), set(["summary","grid","srcasm"]) ),
( query("Machine Clears"), set(["summary","grid","srcasm"]) ) ] ) ] ),
( "Unfilled Pipeline Slots", set(["summary","grid"]),
[ ( query("Back-end Bound Pipeline Slots"), set(["summary","grid","srcasm","collapsed"]),
[ ( "Memory Latency", set(["summary","grid"]),
[ ( query("LLC Miss"), set(["summary","grid","srcasm"]) ),
( query("LLC Hit"), set(["summary","grid","srcasm"]) ),
( query("DTLB Overhead"), set(["summary","grid","srcasm"]) ),
( query("Contested Accesses"), set(["summary","grid","srcasm"]) ),
( query("Data Sharing"), set(["summary","grid","srcasm"]) ) ] ),
( "Memory Replacements", set(["summary","grid"]),
[ ( query("L1D Replacement Percentage"), set(["summary","grid","srcasm"])),
( query("L2Replacement Percentage"), set(["summary","grid","srcasm"])),
( query("LLC Replacement Percentage"), set(["summary","grid","srcasm"])) ] ),
( "Memory Reissues", set(["summary","grid"]),
[ ( query("Loads Blocked by Store Forwarding"), set(["summary","grid","srcasm"])),
( query("Split Loads"), set(["summary","grid","srcasm"])),
( query("Split Stores"), set(["summary","grid","srcasm"])),
( query("4K Aliasing"), set(["summary","grid","srcasm"])) ] ),
( query("DIV Active"), set(["summary","grid","srcasm"])),
( query("Flags Merge Stalls"), set(["summary","grid","srcasm"])),
( query("Slow LEA Stalls"), set(["summary","grid","srcasm"])) ] ),
( query("Front-end Bound Pipeline Slots"), set(["summary","grid","srcasm","collapsed"]),
[ ( query("ICache Misses"), set(["summary","grid","srcasm"])),
( query("ITLB Overhead"), set(["summary","grid","srcasm"])),
( query("DSB to MITE Switch Cost"), set(["summary","grid","srcasm"])) ] ) ] )
]In the first line, an instance variable of the class analysis is created and assigned the ID SNB General Exploration. If an analysis with that ID already exists, then it is extended, not overwritten, just as with metrics.