tools: added frontend_bound metric #8
Conversation
tools: removed the debug statement
…awar/iodlr into add_frontend_bound_metric
tools: removed the debug statement
…awar/iodlr into add_frontend_bound_metric
| SCRIPTS_DIR=`dirname $0` | ||
| source ${SCRIPTS_DIR}/utils.sh | ||
|
|
||
| function init_frontend_bound() { |
There was a problem hiding this comment.
Is this useful/used?
Also, I think it can be reduced to the following, due to the fact that the array is already hardcoded there and there's nothing preventing us from hardcoding it as a string (we could even drop the variable declaration, though it makes the code slightly more readable):
function init_frontend_bound() {
local local_pmus="cpu_clk_unhalted.thread_any,idq_uops_not_delivered.core"
echo $local_pmus
}
| echo " b=idq_uops_not_delivered.core" | ||
| echo "=================================================" | ||
|
|
||
| local a=`return_pmu_value "cpu_clk_unhalted.thread_any" $perf_data_file ` |
There was a problem hiding this comment.
Extra space after $perf_data_file
There was a problem hiding this comment.
Thanks @uttampawar for the PR.
Here are some comments:
- What if HT is not ON? Will threads be correct? You may want to get this programatically.
- Formulas are available on https://download.01.org/perfmon/TMA_Metrics.xlsx
FE BOUND = IDQ_UOPS_NOT_DELIVERED.CORE / SLOTS SLOTS = #Pipeline_Width * CORE_CLKS CORE_CLKS = CPU_CLK_UNHALTED.THREAD
- In future we want to add some checks about the architecture. Some other formulas work only on Skylake and beyond. This one is fine.
--
I realize it yesterday while working on this script. This needs to be addressed for sure one for getting up to date architecture information, exact counter names and formulas to getting various metrics.
There was a problem hiding this comment.
Thanks @uttampawar for the PR.
Here are some comments:
-
What if HT is not ON? Will threads be correct? You may want to get this programatically.
-
Formulas are available on https://download.01.org/perfmon/TMA_Metrics.xlsx
FE BOUND = IDQ_UOPS_NOT_DELIVERED.CORE / SLOTS
SLOTS = #Pipeline_Width * CORE_CLKS
CORE_CLKS = CPU_CLK_UNHALTED.THREAD
- In future we want to add some checks about the architecture. Some other formulas work only on Skylake and beyond. This one is fine.
--
|
General comment is that this was a quick script I needed for my data collection. |
uttampawar
requested review from
catalin-manciu
and removed request for
catalin-manciu
|
@uttampawar is this PR still relevant? |
Hi,
I've added frontend_bound metric calculation script to give high level data about a workload.
@suresh-srinivas Could you review the PR and verify if it gives correct results please? I've copied the formula from edp/skx xml file. I appreciate any comments you may have.