Bad Performance on first Power.From() call in .netCore

[JeCodeFu]

Description
The very first call on Power.From takes up to several seconds.
All the following calls are fine (regarding the performance).

Environment

• .net Core 3.1
• 32bit ARM Platform
• Embedded Linux

Repro/Code
var result = Power.From(5, PowerUnit.Watt);

Expected behavior
The call should return within milliseconds, but it takes up to more than 10 seconds.

Additional context
I Experienced this issue after updating from Version 4.72.0 to Version 4.144.0.
It did not happen in the older version.

[angularsen]

Hi, this is a known issue: #965

Due to a large number of types and type members for all our 100+ quantities and 1000+ units, the JIT spends a long time on the initial usages.

We don't currently have a solution to this, but a few ideas have been floated:

• Break up into core nuget + specific quantity nugets. Only add what you need.
• Switch from struct to class to make use of inheritance
• Make use of extension methods more to avoid duplicating methods for every N quantities
• .NET 7 brings AOT, maybe applications can make use of that to speed up JIT?

[Jens88]

Hi
Thank you for the answer and the ideas how to deal with this.
I did not read all the release notes between V4.72.0 and V4.144.0.
But since you mentioned the large number of types and type members is the causing this issue;

Did you add a lot of new types after V4.72.0?

I know i run on a "low power" cpu and thus i cannot compare my time measurement with all the benchmark tests done listed in issue #965. But the difference between the two versions i mentioned is a factor of 100 times slower (or even more).
So this makes me think that there is yet another problem, unless you did add hundreds of types after V4.72.

[angularsen]

@Jens88 I'm not sure, we added a bunch of units for sure, but 100x sounds a lot.

Are you able to profile the difference between the two? Maybe find what exact hotspot in the code differs between them. Maybe we have introduced some regression.

There was one known performance regression, but it should have been fixed in 4.142.0.
#1069 (comment)

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[angularsen]

@JeCodeFu Long overdue, but I think this improved a lot in v5, plus some recent improvements in #1210.

Could you give it another go?

I can't replicate the extreme slowness you reported though, since you were testing on ARM32 and I'm testing on a fast computer:

The very first call on Power.From takes up to several seconds.

Maybe the application complexity and size also affects how much work JIT needs to do for UnitsNet.

Benchmark

Comparing 4.144.0 with latest v5.

Using TimeItSharp to run a net7.0 console app with this code:

Console.WriteLine(Power.From(5, PowerUnit.Watt));

Results

Before: 850 ms
After: 523 ms

This includes the overhead of starting and running the console app. The idea was to capture any slowness in the JIT.

Before (v4.144.0)

For some reason this crashed the timeit tool, so I had to run fewer iterations on this nuget.

(ins)$ dotnet timeit perf_before.json
TimeIt (v. 0.0.8.0) by Tony Redondo

Warmup count: 5
Count: 10
Number of Scenarios: 1
Exporters: ConsoleExporter, JsonExporter, Datadog

Scenario: Default
  Warming up .....
    Duration: 4,2442191s
  Run ..........
    Duration: 8,4839588s

Name	Mean	StdDev	StdErr	Min	Max	P95	P90	Outliers
Default	843,5567ms	5,356ms	1,6937ms	835,3823ms	850,371ms	850,371ms	849,9721ms	0
├>process.internal_duration_ms	800,9216	4,920471	1,55599	794,7264	809,472	809,472	808,57088	0
├>process.time_to_end_ms	16,29339	1,85835	0,587662	14,0489	18,4476	18,4476	18,429853	0
├>process.time_to_start_ms	26,34177	0,316539	0,100098	25,8493	26,8823	26,8823	26,781357	0
├>runtime.dotnet.cpu.percent	4,483375	0,18614	0,058862	4,246667	4,766667	4,766667	4,734889	0
├>runtime.dotnet.cpu.system	203,385417	47,298275	14,957028	125	270,833333	270,833333	257,465278	0
├>runtime.dotnet.cpu.user	874,348958	35,224503	11,138966	820,3125	916,666667	916,666667	916,666667	0
├>runtime.dotnet.mem.committed	17316024,32	222527,309147	70369,313849	17165516,8	17899008	17899008	17734553,6	0
├>runtime.dotnet.threads.contention_count	0	0	0	0	0	0	0	0
├>runtime.dotnet.threads.contention_time	0	0	0	0	0	0	0	0
├>runtime.dotnet.threads.count	12,8025	0,00527	0,001667	12,8	12,8125	12,8125	12,8125	0
└>runtime.dotnet.threads.workers_count	1,8725	0,004025	0,001273	1,866667	1,875	1,875	1,875	0

After (v5)

(ins)$ dotnet timeit perf.json
TimeIt (v. 0.0.8.0) by Tony Redondo

Warmup count: 10
Count: 20
Number of Scenarios: 1
Exporters: ConsoleExporter, JsonExporter, Datadog

Scenario: Default
  Warming up ..........
    Duration: 5,3425719s
  Run ....................
    Duration: 10,5619372s

Name	Mean	StdDev	StdErr	Min	Max	P95	P90	Outliers
Default	523,2774ms	3,2313ms	0,7413ms	516,9314ms	530,2838ms	529,344ms	526,9549ms	1
├>process.internal_duration_ms	482,944	3,098095	0,692755	476,3648	488,96	488,81664	487,461547	0
├>process.time_to_end_ms	14,252445	0,355083	0,079399	13,5538	14,7657	14,76206	14,751633	0
├>process.time_to_start_ms	26,380142	0,389723	0,089408	25,8933	27,1318	27,09361	26,988073	0
├>runtime.dotnet.cpu.percent	4,203333	0,317501	0,070995	3,755556	4,766667	4,766667	4,713704	0
├>runtime.dotnet.cpu.system	161,458333	77,457144	17,319944	34,722222	312,5	300,347222	277,777778	0
├>runtime.dotnet.cpu.user	848,958333	66,145126	14,7905	694,444444	937,5	937,5	924,768519	0
├>runtime.dotnet.mem.committed	17302459,733333	115220,946958	25764,186983	17084871,111111	17476266,666667	17454603,377778	17412702,814815	0
├>runtime.dotnet.threads.contention_count	0	0	0	0	0	0	0	0
├>runtime.dotnet.threads.contention_time	0	0	0	0	0	0	0	0
├>runtime.dotnet.threads.count	12,666667	0	0	12,666667	12,666667	12,666667	12,666667	0
└>runtime.dotnet.threads.workers_count	1,777778	0	0	1,777778	1,777778	1,777778	1,777778	0

[angularsen]

@tmilnthorp Some benchmarks posted above, can't spot much different in v5 versions in this simple test setup, but significantly faster than v4.

[Jens88]

Hi Andreas

Thanks for the update on this issue. Since we went back in version we had no actual suffer anymore.

I quickly tested the newest version. Unfortunately all the "Undefined" entities vanished, so my code needs some adaption to get back running again. I will come back to this thread when I converted the code and have a test setup again.

[Jens88]

So now the measurement of the very first call of Power.From() call:

V4.72 : 633ms
V4.144: 17256ms (this was the version i reported here)
V5.21 : 9121 ms

So to summarize:

• Yes there is an improvement from Version 4.144
• But the issue is not solved yet, since compared with the V4.72 there is still an increase of 8'488ms (factor 15 slower)