RFR: 8316813: NMT: Using WhiteBox API, virtual memory tracking should also be stressed in JMH tests [v2]

[Afshin Zafari]

On Tue, 13 Feb 2024 07:56:14 GMT, Thomas Stuefe <stuefe at openjdk.org> wrote:

>>> I wonder, though, how much value this benchmark really has. Since we now measure both the mmap call and the registration with NMT. The mmap call may dominate, and/or bring a lot of randomness, since it depends on the context (e.g. do we swap, do we have pages in cache etc).
>> 
>> I think the more data we have the better. Once we start getting more comfortable in this area (we are still learning) we will surely know how to read and use the results, but for now, I personally think that we should be gathering the data anywhere we can, to look for patterns and to learn from them. I'm curious to see, for example, if the randomness you are predicting will show up, and if so, how strong ...
>> 
>> And we can always try to improve it further when we verify that it really suffers from the issue that you see. For example we might want to add what context this code is running in, if we can tell whether we are swapping or not.
>
>> > I wonder, though, how much value this benchmark really has. Since we now measure both the mmap call and the registration with NMT. The mmap call may dominate, and/or bring a lot of randomness, since it depends on the context (e.g. do we swap, do we have pages in cache etc).
>> 
>> I think the more data we have the better. Once we start getting more comfortable in this area (we are still learning) we will surely know how to read and use the results, but for now, I personally think that we should be gathering the data anywhere we can, to look for patterns and to learn from them. I'm curious to see, for example, if the randomness you are predicting will show up, and if so, how strong ...
>> 
>> And we can always try to improve it further when we verify that it really suffers from the issue that you see. For example we might want to add what context this code is running in, if we can tell whether we are swapping or not.
> 
> Okay.
> 
> One thing to try would be to run the benchmark with NMT enabled and disabled. Disabled numbers should be near the mmap cost themselves.

Thank you @tstuefe, @gerard-ziemski and @jdksjolen for your comments.

-------------

PR Comment: https://git.openjdk.org/jdk/pull/17582#issuecomment-1947948613