Perf: double.ToString(CultureInfo) 7x slower on Linux

[tarekgh]

@ianhays commented on Tue Oct 13 2015

test name	date ran	Windows time	Linux time	linux/windows
System.Globalization.Tests.Perf_NumberCultureInfo.ToString(culturestring: "fr")	10/13/2015	3.780472576	29.0264	7.677981896
System.Globalization.Tests.Perf_NumberCultureInfo.ToString(culturestring: "da")	10/13/2015	3.757379192	28.05800001	7.467439025
System.Globalization.Tests.Perf_NumberCultureInfo.ToString(culturestring: "ja")	10/13/2015	3.773344988	27.9326	7.402609643
System.Globalization.Tests.Perf_NumberCultureInfo.ToString(culturestring: "")	10/13/2015	3.806987203	27.96	7.34439033

Perf Test:

[Benchmark]
[InlineData("fr")]
[InlineData("da")]
[InlineData("ja")]
[InlineData("")]
public void ToString(string culturestring)
{
    double number = 104234.343;
    CultureInfo cultureInfo = new CultureInfo(culturestring);
    foreach (var iteration in Benchmark.Iterations)
        using (iteration.StartMeasurement())
        {
            for (int i = 0; i < innerIterations; i++)
            {
                number.ToString(cultureInfo); number.ToString(cultureInfo); number.ToString(cultureInfo);
                number.ToString(cultureInfo); number.ToString(cultureInfo); number.ToString(cultureInfo);
                number.ToString(cultureInfo); number.ToString(cultureInfo); number.ToString(cultureInfo);
            }
        }
}

Although it's an admittedly poor comparative data point, the ToString(CultureInfo) method for DateTime was 1:1 between Linux:Windows so not all ToString methods are equally slow when using CultureInfo. As such, I am unsure of whom to to notify; @steveharter @stephentoub?

I have not yet added equivalent ToString tests to other classes, but can if desired for more data points.

---

@steveharter commented on Mon Nov 16 2015

@ianhays is 7x on windows with coreclr, or desktop clr? In a similar test, I'm getting:
coreclr+windows 4.3x faster than coreclr+linux
desktopclr+windows 18x-25x faster than coreclr+linux
This includes using a simple cpu-only baseline looping test used to normalize the results due to different cpu speeds and\or differences in clr jitting\etc.

---

@ianhays commented on Mon Dec 26 2016

It's interesting that you're seeing such a big difference between Windows-CoreCLR and Windows-Desktop. For my tests they were nearly 1:1. Could you post some numerical results and we can compare data?

I'm running on Ubuntu 14.04 (running in HyperV), CentOS 7.1 (running in HyperV), and Windows 10 running on coreclr. Same computer. Exact test listed above. Using binaries from
\cpvsbuild\Drops\dev14\ProjectK\raw\23513.00\binaries.amd64ret.

CentOS results:

<collection total="4" passed="4" failed="0" skipped="0" name="Test collection for System.Globalization.Tests.Perf_NumberCultureInfo" time="4.209">
  <test name="System.Globalization.Tests.Perf_NumberCultureInfo.ToString(culturestring: &quot;fr&quot;)" type="System.Globalization.Tests.Perf_NumberCultureInfo" method="ToString" time="1.0664355" result="Pass">
    <traits>
      <trait name="Benchmark" value="true" />
    </traits>
    <performance runid="System.Globalization.Tests.dll-CentOS" etl="/home/ianha/rel/results/System.Globalization.Tests.dll-CentOS.csv">
      <metrics>
        <Duration displayName="Duration" unit="msec" />
      </metrics>
      <iterations>
        <iteration index="0" Duration="29.394985914230347" />
        <iteration index="1" Duration="27.538418054580688" />
        <iteration index="2" Duration="28.897694945335388" />
        <iteration index="3" Duration="33.568613886833191" />
        <iteration index="4" Duration="28.437203049659729" />
        <iteration index="5" Duration="27.611016988754272" />
        <iteration index="6" Duration="29.217790007591248" />
        <iteration index="7" Duration="28.009710907936096" />
        <iteration index="8" Duration="28.690199017524719" />
        <iteration index="9" Duration="31.641247987747192" />
        <iteration index="10" Duration="29.398486018180847" />
        <iteration index="11" Duration="27.699916005134583" />
        <iteration index="12" Duration="27.58421802520752" />
        <iteration index="13" Duration="27.673416018486023" />
        <iteration index="14" Duration="28.14900803565979" />
        <iteration index="15" Duration="27.921112060546875" />
        <iteration index="16" Duration="27.766514897346497" />
        <iteration index="17" Duration="27.707115888595581" />
        <iteration index="18" Duration="28.539501070976257" />
        <iteration index="19" Duration="27.511018991470337" />
        <iteration index="20" Duration="27.805315017700195" />
        <iteration index="21" Duration="27.854113936424255" />
        <iteration index="22" Duration="27.897012948989868" />
        <iteration index="23" Duration="31.014458060264587" />
        <iteration index="24" Duration="28.17520797252655" />
        <iteration index="25" Duration="28.119809031486511" />
        <iteration index="26" Duration="28.370004057884216" />
        <iteration index="27" Duration="27.953910946846008" />
        <iteration index="28" Duration="28.936293959617615" />
        <iteration index="29" Duration="29.012992978096008" />
        <iteration index="30" Duration="28.124009013175964" />
        <iteration index="31" Duration="27.519919991493225" />
        <iteration index="32" Duration="27.68511700630188" />
        <iteration index="33" Duration="27.905812978744507" />
        <iteration index="34" Duration="27.937612056732178" />
        <iteration index="35" Duration="27.705316066741943" />
      </iterations>
    </performance>
  </test>
  <test name="System.Globalization.Tests.Perf_NumberCultureInfo.ToString(culturestring: &quot;da&quot;)" type="System.Globalization.Tests.Perf_NumberCultureInfo" method="ToString" time="1.0456405" result="Pass">
    <traits>
      <trait name="Benchmark" value="true" />
    </traits>
    <performance runid="System.Globalization.Tests.dll-CentOS" etl="/home/ianha/rel/results/System.Globalization.Tests.dll-CentOS.csv">
      <metrics>
        <Duration displayName="Duration" unit="msec" />
      </metrics>
      <iterations>
        <iteration index="0" Duration="36.094968914985657" />
        <iteration index="1" Duration="27.743814945220947" />
        <iteration index="2" Duration="27.875913023948669" />
        <iteration index="3" Duration="27.602617979049683" />
        <iteration index="4" Duration="27.816213965415955" />
        <iteration index="5" Duration="28.792296886444092" />
        <iteration index="6" Duration="27.533019065856934" />
        <iteration index="7" Duration="28.085710048675537" />
        <iteration index="8" Duration="28.151907920837402" />
        <iteration index="9" Duration="27.482020020484924" />
        <iteration index="10" Duration="27.996210932731628" />
        <iteration index="11" Duration="28.618600010871887" />
        <iteration index="12" Duration="27.751914978027344" />
        <iteration index="13" Duration="27.638016939163208" />
        <iteration index="14" Duration="27.872413039207458" />
        <iteration index="15" Duration="30.500468015670776" />
        <iteration index="16" Duration="27.872312903404236" />
        <iteration index="17" Duration="27.939612030982971" />
        <iteration index="18" Duration="27.825913071632385" />
        <iteration index="19" Duration="27.495819926261902" />
        <iteration index="20" Duration="27.704716086387634" />
        <iteration index="21" Duration="27.9049129486084" />
        <iteration index="22" Duration="27.931511998176575" />
        <iteration index="23" Duration="27.667815923690796" />
        <iteration index="24" Duration="27.3945209980011" />
        <iteration index="25" Duration="27.917912006378174" />
        <iteration index="26" Duration="27.716416001319885" />
        <iteration index="27" Duration="27.933212041854858" />
        <iteration index="28" Duration="27.486520051956177" />
        <iteration index="29" Duration="29.66008198261261" />
        <iteration index="30" Duration="29.410285949707031" />
        <iteration index="31" Duration="28.070809960365295" />
        <iteration index="32" Duration="27.493919968605042" />
        <iteration index="33" Duration="28.041810035705566" />
        <iteration index="34" Duration="28.174207925796509" />
        <iteration index="35" Duration="27.583817958831787" />
        <iteration index="36" Duration="27.807613968849182" />
      </iterations>
    </performance>
  </test>
  <test name="System.Globalization.Tests.Perf_NumberCultureInfo.ToString(culturestring: &quot;ja&quot;)" type="System.Globalization.Tests.Perf_NumberCultureInfo" method="ToString" time="1.0497453" result="Pass">
    <traits>
      <trait name="Benchmark" value="true" />
    </traits>
    <performance runid="System.Globalization.Tests.dll-CentOS" etl="/home/ianha/rel/results/System.Globalization.Tests.dll-CentOS.csv">
      <metrics>
        <Duration displayName="Duration" unit="msec" />
      </metrics>
      <iterations>
        <iteration index="0" Duration="28.615100026130676" />
        <iteration index="1" Duration="27.676715970039368" />
        <iteration index="2" Duration="27.601817011833191" />
        <iteration index="3" Duration="27.763615012168884" />
        <iteration index="4" Duration="28.683698058128357" />
        <iteration index="5" Duration="27.865113019943237" />
        <iteration index="6" Duration="27.877612948417664" />
        <iteration index="7" Duration="27.683015942573547" />
        <iteration index="8" Duration="28.065410017967224" />
        <iteration index="9" Duration="27.664615988731384" />
        <iteration index="10" Duration="36.985154032707214" />
        <iteration index="11" Duration="30.639065027236938" />
        <iteration index="12" Duration="27.659315943717957" />
        <iteration index="13" Duration="27.68911600112915" />
        <iteration index="14" Duration="27.640017032623291" />
        <iteration index="15" Duration="27.554818987846375" />
        <iteration index="16" Duration="27.4220210313797" />
        <iteration index="17" Duration="27.686416029930115" />
        <iteration index="18" Duration="27.96251106262207" />
        <iteration index="19" Duration="27.818414092063904" />
        <iteration index="20" Duration="28.181107044219971" />
        <iteration index="21" Duration="30.207971930503845" />
        <iteration index="22" Duration="27.66921591758728" />
        <iteration index="23" Duration="27.575117945671082" />
        <iteration index="24" Duration="27.539119005203247" />
        <iteration index="25" Duration="27.812513947486877" />
        <iteration index="26" Duration="28.019811034202576" />
        <iteration index="27" Duration="27.516318917274475" />
        <iteration index="28" Duration="27.816313982009888" />
        <iteration index="29" Duration="27.987210988998413" />
        <iteration index="30" Duration="28.7555969953537" />
        <iteration index="31" Duration="27.774914979934692" />
        <iteration index="32" Duration="28.356604933738708" />
        <iteration index="33" Duration="28.666599035263062" />
        <iteration index="34" Duration="27.587418079376221" />
        <iteration index="35" Duration="30.257572054862976" />
        <iteration index="36" Duration="28.29650604724884" />
      </iterations>
    </performance>
  </test>
  <test name="System.Globalization.Tests.Perf_NumberCultureInfo.ToString(culturestring: &quot;&quot;)" type="System.Globalization.Tests.Perf_NumberCultureInfo" method="ToString" time="1.0471872" result="Pass">
    <traits>
      <trait name="Benchmark" value="true" />
    </traits>
    <performance runid="System.Globalization.Tests.dll-CentOS" etl="/home/ianha/rel/results/System.Globalization.Tests.dll-CentOS.csv">
      <metrics>
        <Duration displayName="Duration" unit="msec" />
      </metrics>
      <iterations>
        <iteration index="0" Duration="28.685899019241333" />
        <iteration index="1" Duration="27.759415030479431" />
        <iteration index="2" Duration="28.962594032287598" />
        <iteration index="3" Duration="27.831114053726196" />
        <iteration index="4" Duration="27.918012022972107" />
        <iteration index="5" Duration="27.727115988731384" />
        <iteration index="6" Duration="27.582617998123169" />
        <iteration index="7" Duration="28.238106966018677" />
        <iteration index="8" Duration="27.707015991210938" />
        <iteration index="9" Duration="27.508718967437744" />
        <iteration index="10" Duration="27.685816049575806" />
        <iteration index="11" Duration="27.990810990333557" />
        <iteration index="12" Duration="36.150267958641052" />
        <iteration index="13" Duration="27.81671404838562" />
        <iteration index="14" Duration="27.862313032150269" />
        <iteration index="15" Duration="28.303506016731262" />
        <iteration index="16" Duration="27.79661500453949" />
        <iteration index="17" Duration="28.096608996391296" />
        <iteration index="18" Duration="27.796614050865173" />
        <iteration index="19" Duration="28.159507989883423" />
        <iteration index="20" Duration="28.012010931968689" />
        <iteration index="21" Duration="27.406520962715149" />
        <iteration index="22" Duration="28.078809022903442" />
        <iteration index="23" Duration="27.770314931869507" />
        <iteration index="24" Duration="27.843314051628113" />
        <iteration index="25" Duration="27.394421100616455" />
        <iteration index="26" Duration="29.00989294052124" />
        <iteration index="27" Duration="30.44566810131073" />
        <iteration index="28" Duration="28.251506090164185" />
        <iteration index="29" Duration="28.132707953453064" />
        <iteration index="30" Duration="28.735197067260742" />
        <iteration index="31" Duration="28.501802086830139" />
        <iteration index="32" Duration="27.959811925888062" />
        <iteration index="33" Duration="27.853813886642456" />
        <iteration index="34" Duration="27.863412976264954" />
        <iteration index="35" Duration="27.652817010879517" />
        <iteration index="36" Duration="27.678416967391968" />
      </iterations>
    </performance>
  </test>
</collection>

Windows 10 on CoreCLR results:

<collection total="4" passed="4" failed="0" skipped="0" name="Test collection for System.Globalization.Tests.Perf_NumberCultureInfo" time="4.087">
  <test name="System.Globalization.Tests.Perf_NumberCultureInfo.ToString(culturestring: &quot;fr&quot;)" type="System.Globalization.Tests.Perf_NumberCultureInfo" method="ToString" time="1.075388" result="Pass">
    <traits>
      <trait name="Benchmark" value="true" />
    </traits>
    <performance runid="System.Globalization.Tests.dll-WindowsCore" etl="D:\tools\testing\relresults\11-16-2015\System.Globalization.Tests.dll-WindowsCore.etl">
      <metrics>
        <Duration displayName="Duration" unit="msec" />
      </metrics>
      <iterations>
        <iteration index="0" Duration="2.1205962986784925" />
        <iteration index="1" Duration="1.544973291548672" />
        <iteration index="2" Duration="1.4910888198560315" />
        <iteration index="3" Duration="1.5059141771470195" />
        <iteration index="4" Duration="1.6136831205324143" />
        <iteration index="5" Duration="1.5338542735803458" />
        <iteration index="6" Duration="1.5224501525871119" />
        <iteration index="7" Duration="1.496505777327684" />
        <iteration index="8" Duration="1.5469690127224567" />
        <iteration index="9" Duration="1.6091214721351434" />
        <iteration index="10" Duration="1.6191000780041804" />
        <iteration index="11" Duration="1.5027780438739455" />
        <iteration index="12" Duration="1.9124710905532538" />
        <iteration index="13" Duration="1.4910888198559178" />
        <iteration index="14" Duration="1.4896633047318346" />
        <iteration index="15" Duration="1.5284373161084659" />
        <iteration index="16" Duration="1.5355648917292228" />
        <iteration index="17" Duration="1.582606890826014" />
        <iteration index="18" Duration="1.4899484077566285" />
        <iteration index="19" Duration="1.4959355712779825" />
        <iteration index="20" Duration="1.5515306611197275" />
        <iteration index="21" Duration="1.5823217878012201" />
        <iteration index="22" Duration="1.4919441289304132" />
        <iteration index="23" Duration="1.6273680657242267" />
        <iteration index="24" Duration="1.496505777327684" />
        <iteration index="25" Duration="1.6883801130376241" />
        <iteration index="26" Duration="1.4970759833772718" />
        <iteration index="27" Duration="1.4885228926324316" />
        <iteration index="28" Duration="1.5184587102395426" />
        <iteration index="29" Duration="1.6456146593133099" />
        <iteration index="30" Duration="1.5187438132643365" />
        <iteration index="31" Duration="1.5190289162891304" />
        <iteration index="32" Duration="1.4899484077566285" />
        <iteration index="33" Duration="1.5090503104200934" />
        <iteration index="34" Duration="1.4967908803524779" />
        <iteration index="35" Duration="1.5107609285691979" />
        <iteration index="36" Duration="1.5928705997198449" />
        <iteration index="37" Duration="1.5743389031059678" />
        <iteration index="38" Duration="1.6444742472139069" />
        <iteration index="39" Duration="1.5680666365597062" />
        <iteration index="40" Duration="1.611402296333722" />
        <iteration index="41" Duration="1.5652156063114262" />
        <iteration index="42" Duration="1.5800409636026416" />
        <iteration index="43" Duration="1.5036333529484409" />
        <iteration index="44" Duration="1.49336964405461" />
        <iteration index="45" Duration="1.4950802622036008" />
        <iteration index="46" Duration="1.5247309767856905" />
        <iteration index="47" Duration="1.489378201706927" />
        <iteration index="48" Duration="1.5746240061308754" />
        <iteration index="49" Duration="1.49622067430289" />
        <iteration index="50" Duration="1.4967908803525916" />
        <iteration index="51" Duration="1.5811813757019308" />
        <iteration index="52" Duration="1.5501051459955306" />
        <iteration index="53" Duration="1.5047737650477302" />
        <iteration index="54" Duration="1.5401265401266073" />
        <iteration index="55" Duration="1.5412669522258966" />
        <iteration index="56" Duration="1.4902335107814224" />
        <iteration index="57" Duration="1.4890930986822468" />
        <iteration index="58" Duration="1.4839612442351608" />
        <iteration index="59" Duration="1.5672113274853245" />
        <iteration index="60" Duration="1.4896633047318346" />
        <iteration index="61" Duration="1.6245170354759466" />
        <iteration index="62" Duration="1.5552370004425029" />
        <iteration index="63" Duration="1.5164629890657579" />
        <iteration index="64" Duration="1.6176745628799836" />
        <iteration index="65" Duration="1.5663560184108292" />
        <iteration index="66" Duration="1.513326855792684" />
        <iteration index="67" Duration="1.6843886706900548" />
        <iteration index="68" Duration="1.8161062681610929" />
        <iteration index="69" Duration="1.5187438132643365" />
        <iteration index="70" Duration="1.5660709153859216" />
        <iteration index="71" Duration="1.4927994380049086" />
        <iteration index="72" Duration="1.5392712310519983" />
        <iteration index="73" Duration="1.5635049881625491" />
        <iteration index="74" Duration="1.5463988066728689" />
        <iteration index="75" Duration="1.5161778860408504" />
        <iteration index="76" Duration="1.4930845410297025" />
        <iteration index="77" Duration="1.5840324059502109" />
        <iteration index="78" Duration="1.4899484077566285" />
        <iteration index="79" Duration="1.5153225769664687" />
        <iteration index="80" Duration="1.6145384296069096" />
        <iteration index="81" Duration="1.5769048303295676" />
        <iteration index="82" Duration="1.5161778860408504" />
        <iteration index="83" Duration="1.5119013406684871" />
        <iteration index="84" Duration="1.5592284427900722" />
        <iteration index="85" Duration="1.4879526865827302" />
        <iteration index="86" Duration="1.5908748785460602" />
        <iteration index="87" Duration="1.5081950013457117" />
        <iteration index="88" Duration="1.6784015071685872" />
        <iteration index="89" Duration="1.8466122918176779" />
        <iteration index="90" Duration="1.7445454089289569" />
        <iteration index="91" Duration="1.564930503286746" />
        <iteration index="92" Duration="1.7029203673039319" />
        <iteration index="93" Duration="1.5272969040092903" />
        <iteration index="94" Duration="1.5264415949347949" />
        <iteration index="95" Duration="1.5663560184108292" />
        <iteration index="96" Duration="1.5592284427900722" />
        <iteration index="97" Duration="1.525301182835392" />
        <iteration index="98" Duration="1.8417655403957269" />
        <iteration index="99" Duration="1.5506753520451184" />
        <iteration index="100" Duration="1.523305461661721" />
      </iterations>
    </performance>
  </test>
  <test name="System.Globalization.Tests.Perf_NumberCultureInfo.ToString(culturestring: &quot;da&quot;)" type="System.Globalization.Tests.Perf_NumberCultureInfo" method="ToString" time="1.0043532" result="Pass">
    <traits>
      <trait name="Benchmark" value="true" />
    </traits>
    <performance runid="System.Globalization.Tests.dll-WindowsCore" etl="D:\tools\testing\relresults\11-16-2015\System.Globalization.Tests.dll-WindowsCore.etl">
      <metrics>
        <Duration displayName="Duration" unit="msec" />
      </metrics>
      <iterations>
        <iteration index="0" Duration="1.8386294071224256" />
        <iteration index="1" Duration="1.6105469872593403" />
        <iteration index="2" Duration="1.6895205251369134" />
        <iteration index="3" Duration="1.5780452424287432" />
        <iteration index="4" Duration="1.4999270136256655" />
        <iteration index="5" Duration="1.4959355712778688" />
        <iteration index="6" Duration="1.4985014985013549" />
        <iteration index="7" Duration="1.708622427800492" />
        <iteration index="8" Duration="1.5258713888852071" />
        <iteration index="9" Duration="1.4979312924517671" />
        <iteration index="10" Duration="1.5375606129030075" />
        <iteration index="11" Duration="1.5429775703748874" />
        <iteration index="12" Duration="1.5999981753407155" />
        <iteration index="13" Duration="1.7941533352493479" />
        <iteration index="14" Duration="1.4959355712780962" />
        <iteration index="15" Duration="1.6005683813903033" />
        <iteration index="16" Duration="1.4950802622036008" />
        <iteration index="17" Duration="1.6176745628799836" />
        <iteration index="18" Duration="1.6279382717739281" />
        <iteration index="19" Duration="1.5831770968757155" />
        <iteration index="20" Duration="1.5150374739416748" />
        <iteration index="21" Duration="1.4927994380050222" />
        <iteration index="22" Duration="1.7225924760170983" />
        <iteration index="23" Duration="1.5463988066728689" />
        <iteration index="24" Duration="1.4922292319552071" />
        <iteration index="25" Duration="1.5418371582754844" />
        <iteration index="26" Duration="1.6863843918638395" />
        <iteration index="27" Duration="1.6866694948887471" />
        <iteration index="28" Duration="1.5093354134451147" />
        <iteration index="29" Duration="1.6019938965143865" />
        <iteration index="30" Duration="1.5332840675305306" />
        <iteration index="31" Duration="1.6079810600358542" />
        <iteration index="32" Duration="1.8936542909145828" />
        <iteration index="33" Duration="1.5150374739414474" />
        <iteration index="34" Duration="1.7682089599898063" />
        <iteration index="35" Duration="1.8825352729461429" />
        <iteration index="36" Duration="1.6347807443696638" />
        <iteration index="37" Duration="1.5888791573722756" />
        <iteration index="38" Duration="1.6230915203518634" />
        <iteration index="39" Duration="1.6667122831506731" />
        <iteration index="40" Duration="1.4987866015262625" />
        <iteration index="41" Duration="1.6812525374170946" />
        <iteration index="42" Duration="1.5592284427900722" />
        <iteration index="43" Duration="1.5680666365597062" />
        <iteration index="44" Duration="1.8534547644137547" />
        <iteration index="45" Duration="1.7160351064460428" />
        <iteration index="46" Duration="1.7049160884776029" />
        <iteration index="47" Duration="1.9942956586792207" />
        <iteration index="48" Duration="1.5170331951155731" />
        <iteration index="49" Duration="1.7921576140754496" />
        <iteration index="50" Duration="1.608551266085442" />
        <iteration index="51" Duration="1.5689219456342016" />
        <iteration index="52" Duration="1.6219511082524605" />
        <iteration index="53" Duration="1.6099767812097525" />
        <iteration index="54" Duration="2.0721287844576182" />
        <iteration index="55" Duration="1.8158211651361853" />
        <iteration index="56" Duration="1.5566625155668135" />
        <iteration index="57" Duration="1.5042035589981424" />
        <iteration index="58" Duration="1.5406967461760814" />
        <iteration index="59" Duration="1.7171755185452184" />
        <iteration index="60" Duration="1.6071257509613588" />
        <iteration index="61" Duration="1.6002832783653957" />
        <iteration index="62" Duration="1.532143655431355" />
        <iteration index="63" Duration="1.6516018228346638" />
        <iteration index="64" Duration="2.1553788677076682" />
        <iteration index="65" Duration="1.7268690213895752" />
        <iteration index="66" Duration="1.6550230591326454" />
        <iteration index="67" Duration="1.5190289162892441" />
        <iteration index="68" Duration="1.5942961148441555" />
        <iteration index="69" Duration="1.5837473029253033" />
        <iteration index="70" Duration="1.8617227521335735" />
        <iteration index="71" Duration="1.68780990698815" />
        <iteration index="72" Duration="1.7074820157013164" />
        <iteration index="73" Duration="1.8349230677997639" />
        <iteration index="74" Duration="1.6838184646403533" />
        <iteration index="75" Duration="1.6661420771010853" />
        <iteration index="76" Duration="1.7034905733537471" />
        <iteration index="77" Duration="1.6695633133988395" />
        <iteration index="78" Duration="1.8697056368289395" />
        <iteration index="79" Duration="1.6803972283423718" />
        <iteration index="80" Duration="1.6977885128569596" />
        <iteration index="81" Duration="1.6652867680263626" />
        <iteration index="82" Duration="1.7111883550239781" />
        <iteration index="83" Duration="1.9603683987245404" />
        <iteration index="84" Duration="1.6732696527217286" />
        <iteration index="85" Duration="1.6752653738956269" />
        <iteration index="86" Duration="1.6635761498775992" />
        <iteration index="87" Duration="1.6983587189067748" />
        <iteration index="88" Duration="1.6709888285229226" />
        <iteration index="89" Duration="1.6912311432859042" />
        <iteration index="90" Duration="1.8292210073032038" />
        <iteration index="91" Duration="1.6735547557464088" />
        <iteration index="92" Duration="1.8876671273933425" />
        <iteration index="93" Duration="1.6071257509613588" />
        <iteration index="94" Duration="1.5734835940315861" />
        <iteration index="95" Duration="1.7231626820669135" />
        <iteration index="96" Duration="1.7476815422021446" />
        <iteration index="97" Duration="1.5626496790880537" />
        <iteration index="98" Duration="1.5706325637834198" />
        <iteration index="99" Duration="1.5863132301487894" />
        <iteration index="100" Duration="1.5894493634218634" />
      </iterations>
    </performance>
  </test>
  <test name="System.Globalization.Tests.Perf_NumberCultureInfo.ToString(culturestring: &quot;ja&quot;)" type="System.Globalization.Tests.Perf_NumberCultureInfo" method="ToString" time="1.0038747" result="Pass">
    <traits>
      <trait name="Benchmark" value="true" />
    </traits>
    <performance runid="System.Globalization.Tests.dll-WindowsCore" etl="D:\tools\testing\relresults\11-16-2015\System.Globalization.Tests.dll-WindowsCore.etl">
      <metrics>
        <Duration displayName="Duration" unit="msec" />
      </metrics>
      <iterations>
        <iteration index="0" Duration="1.6245170354759466" />
        <iteration index="1" Duration="1.5646454002617247" />
        <iteration index="2" Duration="1.5215948435129576" />
        <iteration index="3" Duration="1.5683517395846138" />
        <iteration index="4" Duration="1.5281522130840131" />
        <iteration index="5" Duration="1.5729133879822257" />
        <iteration index="6" Duration="1.6721292406223256" />
        <iteration index="7" Duration="1.6216660052277803" />
        <iteration index="8" Duration="1.6872397009383349" />
        <iteration index="9" Duration="1.6838184646403533" />
        <iteration index="10" Duration="1.7280094334892055" />
        <iteration index="11" Duration="1.6678526952500761" />
        <iteration index="12" Duration="1.6344956413449836" />
        <iteration index="13" Duration="1.7647877236918248" />
        <iteration index="14" Duration="1.73171577281164" />
        <iteration index="15" Duration="2.0265123004846828" />
        <iteration index="16" Duration="1.7023501612543441" />
        <iteration index="17" Duration="1.7990000866711853" />
        <iteration index="18" Duration="1.8953649090635736" />
        <iteration index="19" Duration="1.9843170528101837" />
        <iteration index="20" Duration="1.7014948521800761" />
        <iteration index="21" Duration="1.8568760007115088" />
        <iteration index="22" Duration="1.8936542909145828" />
        <iteration index="23" Duration="1.8309316254521946" />
        <iteration index="24" Duration="1.7012097491547138" />
        <iteration index="25" Duration="2.1180303714550064" />
        <iteration index="26" Duration="1.84290595249513" />
        <iteration index="27" Duration="1.9027775877093518" />
        <iteration index="28" Duration="2.02622719746023" />
        <iteration index="29" Duration="1.9769043741644055" />
        <iteration index="30" Duration="1.7901618929013239" />
        <iteration index="31" Duration="1.7539538087480651" />
        <iteration index="32" Duration="1.9971466889278418" />
        <iteration index="33" Duration="1.6054151328125954" />
        <iteration index="34" Duration="1.6362062594939744" />
        <iteration index="35" Duration="1.6812525374170946" />
        <iteration index="36" Duration="1.7796130809829265" />
        <iteration index="37" Duration="1.7556644268975106" />
        <iteration index="38" Duration="1.6644314589520945" />
        <iteration index="39" Duration="1.5555221034669557" />
        <iteration index="40" Duration="1.727154124414028" />
        <iteration index="41" Duration="1.5458286006232811" />
        <iteration index="42" Duration="1.5714878728576878" />
        <iteration index="43" Duration="1.8409102313212315" />
        <iteration index="44" Duration="1.8332124496510005" />
        <iteration index="45" Duration="1.7559495299224182" />
        <iteration index="46" Duration="1.5398414371015861" />
        <iteration index="47" Duration="1.6145384296069096" />
        <iteration index="48" Duration="1.7094777368752148" />
        <iteration index="49" Duration="1.6592996045051223" />
        <iteration index="50" Duration="1.6815376404420022" />
        <iteration index="51" Duration="1.5999981753407155" />
        <iteration index="52" Duration="1.61026188423466" />
        <iteration index="53" Duration="1.586598333173697" />
        <iteration index="54" Duration="1.64875079258627" />
        <iteration index="55" Duration="1.6481805865364549" />
        <iteration index="56" Duration="1.5444030854987432" />
        <iteration index="57" Duration="1.538986128027318" />
        <iteration index="58" Duration="2.1610809282042283" />
        <iteration index="59" Duration="1.7188861366944366" />
        <iteration index="60" Duration="1.8232338437819635" />
        <iteration index="61" Duration="1.653312440983882" />
        <iteration index="62" Duration="1.5592284427898448" />
        <iteration index="63" Duration="1.5529561762441517" />
        <iteration index="64" Duration="1.562934782112734" />
        <iteration index="65" Duration="1.5903046724965861" />
        <iteration index="66" Duration="1.5358499947542441" />
        <iteration index="67" Duration="1.6937970705093903" />
        <iteration index="68" Duration="1.6476103804870945" />
        <iteration index="69" Duration="1.6396274957919559" />
        <iteration index="70" Duration="1.5024929408491516" />
        <iteration index="71" Duration="1.6744100648211315" />
        <iteration index="72" Duration="1.5449732915485583" />
        <iteration index="73" Duration="1.6980736158816399" />
        <iteration index="74" Duration="1.6219511082526878" />
        <iteration index="75" Duration="1.5452583945734659" />
        <iteration index="76" Duration="1.5272969040088356" />
        <iteration index="77" Duration="1.6076959570109466" />
        <iteration index="78" Duration="1.6915162463110391" />
        <iteration index="79" Duration="1.6359211564690668" />
        <iteration index="80" Duration="1.5270118009843827" />
        <iteration index="81" Duration="2.0008530282502761" />
        <iteration index="82" Duration="1.7191712397193442" />
        <iteration index="83" Duration="1.5498200429710778" />
        <iteration index="84" Duration="1.7984298806218249" />
        <iteration index="85" Duration="1.6516018228344365" />
        <iteration index="86" Duration="1.5506753520453458" />
        <iteration index="87" Duration="1.8771183154744904" />
        <iteration index="88" Duration="1.8685652247295366" />
        <iteration index="89" Duration="1.5948663208937433" />
        <iteration index="90" Duration="1.7077671187257693" />
        <iteration index="91" Duration="1.5093354134451147" />
        <iteration index="92" Duration="1.5307181403068171" />
        <iteration index="93" Duration="1.5002121166508005" />
        <iteration index="94" Duration="1.5441179824742903" />
        <iteration index="95" Duration="1.627368065724113" />
        <iteration index="96" Duration="1.6644314589520945" />
        <iteration index="97" Duration="2.0273676095594055" />
        <iteration index="98" Duration="1.7014948521796214" />
        <iteration index="99" Duration="1.8865267152937122" />
        <iteration index="100" Duration="1.5284373161084659" />
      </iterations>
    </performance>
  </test>
  <test name="System.Globalization.Tests.Perf_NumberCultureInfo.ToString(culturestring: &quot;&quot;)" type="System.Globalization.Tests.Perf_NumberCultureInfo" method="ToString" time="1.0029813" result="Pass">
    <traits>
      <trait name="Benchmark" value="true" />
    </traits>
    <performance runid="System.Globalization.Tests.dll-WindowsCore" etl="D:\tools\testing\relresults\11-16-2015\System.Globalization.Tests.dll-WindowsCore.etl">
      <metrics>
        <Duration displayName="Duration" unit="msec" />
      </metrics>
      <iterations>
        <iteration index="0" Duration="1.6849588767395289" />
        <iteration index="1" Duration="1.4982163954769021" />
        <iteration index="2" Duration="1.6393423927670483" />
        <iteration index="3" Duration="1.6955076886583811" />
        <iteration index="4" Duration="1.4999270136258929" />
        <iteration index="5" Duration="1.542122261300392" />
        <iteration index="6" Duration="1.4982163954764474" />
        <iteration index="7" Duration="1.4982163954764474" />
        <iteration index="8" Duration="1.9603683987247678" />
        <iteration index="9" Duration="2.1274387712742282" />
        <iteration index="10" Duration="1.821523225632518" />
        <iteration index="11" Duration="1.5908748785459466" />
        <iteration index="12" Duration="1.5706325637834198" />
        <iteration index="13" Duration="1.5683517395846138" />
        <iteration index="14" Duration="1.7334263909606307" />
        <iteration index="15" Duration="1.5478243217967247" />
        <iteration index="16" Duration="1.7482517482517324" />
        <iteration index="17" Duration="1.5338542735803458" />
        <iteration index="18" Duration="1.5107609285691979" />
        <iteration index="19" Duration="1.5635049881625491" />
        <iteration index="20" Duration="1.5894493634218634" />
        <iteration index="21" Duration="1.5270118009843827" />
        <iteration index="22" Duration="1.5250160798109391" />
        <iteration index="23" Duration="1.5076247952956692" />
        <iteration index="24" Duration="2.0099763250450451" />
        <iteration index="25" Duration="1.6276531687490206" />
        <iteration index="26" Duration="1.5808962726773643" />
        <iteration index="27" Duration="1.6424785260401222" />
        <iteration index="28" Duration="1.5287224191333735" />
        <iteration index="29" Duration="1.6963629977326491" />
        <iteration index="30" Duration="1.5284373161084659" />
        <iteration index="31" Duration="1.5968620420671868" />
        <iteration index="32" Duration="1.5903046724961314" />
        <iteration index="33" Duration="1.5666411214356231" />
        <iteration index="34" Duration="1.7379880393577878" />
        <iteration index="35" Duration="1.5190289162892441" />
        <iteration index="36" Duration="1.5492498369208079" />
        <iteration index="37" Duration="1.5572327216159465" />
        <iteration index="38" Duration="1.5954365269431037" />
        <iteration index="39" Duration="1.6792568162431962" />
        <iteration index="40" Duration="1.6245170354759466" />
        <iteration index="41" Duration="1.6937970705093903" />
        <iteration index="42" Duration="1.8363485829236197" />
        <iteration index="43" Duration="1.9831766407110081" />
        <iteration index="44" Duration="1.7331412879361778" />
        <iteration index="45" Duration="1.7493921603509079" />
        <iteration index="46" Duration="2.2497479689259308" />
        <iteration index="47" Duration="1.801566013894444" />
        <iteration index="48" Duration="1.7827492142564552" />
        <iteration index="49" Duration="1.7029203673041593" />
        <iteration index="50" Duration="1.6216660052277803" />
        <iteration index="51" Duration="1.6296488899229189" />
        <iteration index="52" Duration="1.7000693370555382" />
        <iteration index="53" Duration="1.522165049562318" />
        <iteration index="54" Duration="1.5712027698327802" />
        <iteration index="55" Duration="2.0142528704172946" />
        <iteration index="56" Duration="1.5680666365597062" />
        <iteration index="57" Duration="1.5572327216164012" />
        <iteration index="58" Duration="1.5954365269435584" />
        <iteration index="59" Duration="1.5446881885241055" />
        <iteration index="60" Duration="1.5806111696524567" />
        <iteration index="61" Duration="1.6772610950692979" />
        <iteration index="62" Duration="1.523020358636586" />
        <iteration index="63" Duration="2.0242314762858769" />
        <iteration index="64" Duration="1.6584442954303995" />
        <iteration index="65" Duration="1.515322576966355" />
        <iteration index="66" Duration="1.5176034011651609" />
        <iteration index="67" Duration="1.5307181403072718" />
        <iteration index="68" Duration="1.6276531687490206" />
        <iteration index="69" Duration="1.6228064173269559" />
        <iteration index="70" Duration="1.6829631555660853" />
        <iteration index="71" Duration="1.6079810600358542" />
        <iteration index="72" Duration="1.7351370091096214" />
        <iteration index="73" Duration="1.6775461980942055" />
        <iteration index="74" Duration="1.650176307710808" />
        <iteration index="75" Duration="1.7619366934436584" />
        <iteration index="76" Duration="1.6878099069876953" />
        <iteration index="77" Duration="1.7867406566037971" />
        <iteration index="78" Duration="1.8086935895157694" />
        <iteration index="79" Duration="1.7000693370555382" />
        <iteration index="80" Duration="1.5828919938512627" />
        <iteration index="81" Duration="1.8468973948424718" />
        <iteration index="82" Duration="1.6228064173269559" />
        <iteration index="83" Duration="1.6467550714123718" />
        <iteration index="84" Duration="1.7733408144367786" />
        <iteration index="85" Duration="1.6547379561079651" />
        <iteration index="86" Duration="1.6855290827893441" />
        <iteration index="87" Duration="1.6433338351143902" />
        <iteration index="88" Duration="1.5828919938508079" />
        <iteration index="89" Duration="1.5700623577336046" />
        <iteration index="90" Duration="1.9985722040514702" />
        <iteration index="91" Duration="1.7610813843693904" />
        <iteration index="92" Duration="1.9201688722237122" />
        <iteration index="93" Duration="1.9116157814787584" />
        <iteration index="94" Duration="1.5891642603969558" />
        <iteration index="95" Duration="1.8862416122688046" />
        <iteration index="96" Duration="1.7191712397193442" />
        <iteration index="97" Duration="1.9652151501463777" />
        <iteration index="98" Duration="1.6541677500586047" />
        <iteration index="99" Duration="1.5985726602166324" />
        <iteration index="100" Duration="1.8788289336234811" />
      </iterations>
    </performance>
  </test>
</collection>

---

@steveharter commented on Mon Nov 16 2015

Please assign to owner of PAL library, as this is not a globalization issue but instead the implementation of the PAL layer. The method _ecvt (\src\pal\src\cruntime\misctls.cpp) is taking up 95%+ of the time.

Note: the benchmarck test should probably "warm up" the current culture by performing a number.ToString(cultureInfo) before the start of the timing. This will force a one-time load of all of the localization information into the cultureinfo (like the character for the thousands separator, etc).

There are a few inefficiencies in _ecvt:

• Very large string buffer to hold the result (~350 chars)
• A copy of the entire string buffer due to the format string using exponent syntax %.348e" where the "e" is almost at the end of string typically with 100s of zeros (for small numbers).
• Re-writing sprintf_s or using an alternate Linux method may be more efficient. One inefficiency is that sprintf_s has a string-based format string which must be parsed

Changing the implementation to not use the large buffer makes the code run 4x+ faster, such as inserting this code at the beginning (sample - won't work in all cases):
pThread = InternalGetCurrentThread();
lpStartOfReturnBuffer = lpReturnBuffer = pThread->crtInfo.ECVTBuffer;
sprintf_s(lpStartOfReturnBuffer, sizeof(TempBuffer), "%lf", value);
*dec = 1;
*sign = 0;
goto done;

---

@ianhays commented on Mon Nov 16 2015

Please assign to owner of PAL library, as this is not a globalization issue but instead the implementation of the PAL layer. The method _ecvt (\src\pal\src\cruntime\misctls.cpp) is taking up 95%+ of the time.

Thanks for looking into this Steve, will do. Know who that is by any chance?

Note: the benchmarck test should probably "warm up" the current culture by performing a number.ToString(cultureInfo) before the start of the timing. This will force a one-time load of all of the localization information into the cultureinfo (like the character for the thousands separator, etc).

This is taken into account in the analysis. The first few iterations are ignored.

---

@stephentoub commented on Tue Nov 17 2015

cc: @sergiy-k

---

@steveharter commented on Tue Nov 17 2015

Not sure who the owner(s) are or the original author. I am willing to take it, but it would have to come after rc2.

Another option is to measure the performance of using ICU to do that parsing.

---

@gkhurana commented on Wed Dec 09 2015

A few alternatives to sprintf_s, ecvt are snprintf and ecvt_r on unix.
ecvt* family of functions are now obsolete and not recommended.

---

@krwq commented on Mon Nov 28 2016

@ianhays Could you please share the units of the measurements and how does that impact the whole framework? I think we should not be optimizing 1ns->40ns regression or really rarely used calls as it is simply not worth the effort. I will leave this open as up-for-grabs. Please change if you believe this should be fixed sooner.

---

@ianhays commented on Mon Dec 05 2016

The measurements were made using msbuild /t:buildandtest /p:Performance=true which calls the xunit.performance runner. It looks like this test has actually been removed since the measurement was made, so there isn't any way to repro it in the same way.

really rarely used calls as it is simply not worth the effort.

I wrote a perf test for this method because usage data showed it high a relatively high utilization rate

I think we should not be optimizing 1ns->40ns regression

If you don't feel it is a necessary optimization please feel free to close out the issue. These micro-measurements were done on thousands of iterations which is dissimilar to the actual usage, so whether it is worth improving is up to you.

---

@AndyAyersMS commented on Mon Dec 05 2016

There are quite a few low-level native helper routines that appear to be slower on linux, for instance much of the math library. See for instance some of the discussion and data in dotnet/coreclr#4847.

---

@karelz commented on Wed Mar 22 2017

The issue seems to be blocking https://github.com/dotnet/corefx/issues/16636 (which is blocking .NET Core adoption for a customer who reached to us via CSS).
@mellinoe can you please help push it forward as part of your perf work in 2.0? (feel free to work with @krwq and @tarekgh, the area owners)
See https://github.com/dotnet/corefx/issues/16636 for additional link to CoreCLR (maybe there are some dupes here).

---

@tannergooding commented on Wed Mar 22 2017

I have https://github.com/dotnet/coreclr/issues/9373, which is tracking the general perf issues with the math library on Linux.

---

@tarekgh commented on Wed Mar 22 2017

Looks to me there is no action from the library side on this one. all changes would be from the runtime side.

@karelz looks there are some tracking issues targeting fixing this issue. so I think we should close this one. I already ensured dotnet/coreclr#4847 and dotnet/coreclr#9373 are marked to v2.0.0

[tarekgh]

@tannergooding I assigned it to you as you already looking at similar issue.

[karelz]

Just to reiterate importance:
The issue seems to be blocking dotnet/corefx#16636 (which is blocking .NET Core adoption for a customer who reached to us via CSS). See benchmatk https://github.com/dotnet/coreclr/issues/5558#issuecomment-288569461 for details

[karelz]

cc @mellinoe to help push it forward

[tannergooding]

We should likely be special casing a few numbers (0, NaN, -Inf, +Inf) and then following the algorithm described here for everything else Printing Floating-Point Numbers Quickly and Accurately. NOTE: Windows follows this algorithm, but additionally continues generating insignificant digits so that powers of two (which are exactly representable) can be printed.

cc @dcwuser, as he generally has helpful feedback in this area as well.

[dcwuser]

Looking at the _ecvt code (in src/pal/src/cruntime/misctls.cpp), it looks to me like this implementation is fundamentally taking the wrong approach. It's essentially calling into sprintf to do the core of its job, then using string munging to fix up the output. That basically means that we do the required work twice: inside sprintf we parse a format string and do the actual digit extraction (for all digits, even if we ultimately only want 2); but inside ToString we have already parsed one format string, and inside _ecvt we are re-doing the extraction of the wanted digits via string manipulation.

The right approach is to go directly from the double to a minimal array of digits inside _ecvt. There should never be a call into sprintf. There should be almost no string munging. (We might be returning a string, but it's best to think of returning an array of digits; if I were writing this for maximum clarity in managed code I would return a List of int.) The approach in the paper that @tannergooding cites looks fantastic and optimal, but even the more naive, plodding approach (multiply by a power to ten to put one digit to the left of the decimal point; pick it off by using floor and casting the result to an int; subtract off the extracted value and multiply by 10 to bring the next digit to the left of the decimal; repeat until the desired number of digits are extracted) is likely to be better than this approach.

It may be possible to get some perf improvement by making small changes to the current approach, and that might even be the right approach to rapidly un-blocking the customer cited by @karelz . But a fundamentally different approach is the correct fix, particularly if that's what's being done on the Windows side.

It might also be worthwhile to look at what's being done inside sprintf when it's handed a double. Presumably it's calling something analogous to _ecvt, and we might be able to call that function directly in our _ecvt implementation.

[jkotas]

_ecvt in the PAL is trying to emulate behavior of Windows _ecvt - in not a very efficient way . The right fix should be to delete _ecvt in the PAL and just implement the DoubleToNumber method directly. Note that NumberToDouble is on that plan already.

[dcwuser]

I found a few ecvt implementations via search engine, and they all do what I describe as the "naive, plodding" approach. No calls to sprintf, no string munging. Since this is part of libc, we should able to call directly into the Linux libc implementation.

https://opensource.apple.com/source/Libc/Libc-167/gen.subproj/ppc.subproj/ecvt.c
https://sourcecodebrowser.com/linux86/0.16.17/ecvt_8c.html#aa9d3fe3843186a88d67641f87f62f712

[tarekgh]

I recall ecvt was not available on MAC but I am not sure. we need to validate it exists

[tarekgh]

now I recall, we cannot use ecvt/fcvt because it is not thread safe as it uses a fixed buffer to return the results for all calls. the safe one ecvt_r is not available on MAC.

so I think what @jkotas suggested is the way to go.

[dcwuser]

@tarekgh, you appear to be correct. Who writes a method like that in a general-use base library?

Forgive me for being dense, but it's not entirely clear to me what you mean, @jkotas. I looked at the code path underneath Double.Parse, and it looks like it feeds into a series of methods that are fully implemented in managed code, without any extern calls. Do you mean, we should do the same for Double.ToString: write a digit extractor fully in managed code -- essentially a managed, well-implemented ecvt function?

[tannergooding]

@dcwuser, I believe he is referring to this line here: https://github.com/dotnet/coreclr/blob/52a816d3011f4d03b80b5940dc036b40d701f52d/src/classlibnative/bcltype/number.cpp#L152 (Note: the function is implemented in raw asm for x86: https://github.com/dotnet/coreclr/blob/32f0f9721afb584b4a14d69135bea7ddc129f755/src/vm/i386/fptext.asm#L78).

I think the quick-fix is to just modify the PAL implementation of _ecvt (this is mostly because of the x86 changes that would be required as well) to be more efficient and the longer term fix is to get rid of _ecvt altogether (and optionally move the implementation entirely to managed code).

[tarekgh]

Do you mean, we should do the same for Double.ToString: write a digit extractor fully in managed code -- essentially a managed, well-implemented ecvt function?

what I meant is we implement DoubleToNumber fully independent of the OS. either in managed if there will not be any concern with the perf or in native side if the managed side perf would not be acceptable.
We can use whatever reasonable algorithm suggested in this thread in the implementation.

[tmds]

Linux (glibc) has ecvt_r, so calling ecvt_r on platforms that have it, should yield a performance improvement.

[tannergooding]

@tarekgh, is the performance on Mac currently a problem/concern as well (I don't see any numbers)?

If not, adding the appropriate conditional so that ecvt_r is called on Linux seems sufficient for the time being.

After going over it a bit, the issue with all the 'sufficiently fast' implementations is that they rely on high-precision integer arithmetic (that is BigInteger math) to make things fast and accurate. Given that there is no BigInteger type built-in, one has to implement a BigInteger implementation that contains, at a minimum, just the required behavior (essentially limited to: left shift, 2^x, 10^x, multiply, and divide).

As such, going the route for the fast implementation probably requires a bit more bake time (to make sure it is good and working correctly) then we likely want to spend for a quick-turnaround.

Now, if we want to look at this a different way, the biggest issue with the current implementation isn't that it is calling sprintf_s. The issue is that it is calling sprintf_s with a format string of %.348e (which prints 348 digits + exponent). This requires us to do additional parsing, rounding, etc... If we instead did %.##f (where ## was the precision requested). We could then do one loop over the resulting string to pull out both dec and sign and then a memcpy to copy the appropriate bits (everything except for the sign and decimal point) into the result buffer. This should be significantly faster and simpler than what we have now.