[erlang-bugs] Erlang 18.0-rc1 is behaving slower than 17.4 (and 17.5)

[Jesper Louis Andersen]

Ok, and I spoke too soon and reported numbers for 17.5 :P

I've collected 17.5 and 18.0-rc1-109-gb5f3973 by the same method as in my
previous post. The summaries in R are:

> r17 <- read.csv('data-17.5.txt', header=FALSE)
> r18 <- read.csv('data_OTP-18.0-rc1-109-gb5f3973.txt', header=FALSE)
> r17$version <- 17
> r18$version <- 18
> x <- rbind(r17, r18)
> summary(r17)
       V1            version
 Min.   :0.1900   Min.   :17
 1st Qu.:0.1930   1st Qu.:17
 Median :0.1940   Median :17
 Mean   :0.1946   Mean   :17
 3rd Qu.:0.1960   3rd Qu.:17
 Max.   :0.2040   Max.   :17
> summary(r18)
       V1            version
 Min.   :0.2100   Min.   :18
 1st Qu.:0.2790   1st Qu.:18
 Median :0.2840   Median :18
 Mean   :0.2968   Mean   :18
 3rd Qu.:0.3493   3rd Qu.:18
 Max.   :0.4050   Max.   :18

This is more like it! They are vastly different, with the r18 version
having a much much larger spread. How large?

install.packages('ggplot2')
m <- ggplot(x, aes(x = V1, colour=version, group=version)
m + geom_density(fill=NA)

The kernel density plot on imgur:

http://imgur.com/9i8YE0U

which clearly shows that bootup of 17.5 is consistent, whereas bootup of
the 18.0-rc1++ version I run is highly erratic. Is this due to timer wheel
bootup and the new time tracking code?


On Mon, Apr 13, 2015 at 10:38 PM, Jesper Louis Andersen <
jesper.louis.andersen@REDACTED> wrote:

>
> On Mon, Apr 13, 2015 at 6:44 PM, José Valim <
> jose.valim@REDACTED> wrote:
>
>> 2. The time to start, write to stdio and shutdown the erlang system is
>> higher on 18.0
>
>
> I was interested in this one, because I think I've perceived startup times
> being less regular, and this is a good check. I did:
>
> [jlouis@REDACTED 2]$ for i in $(seq 1 100); do time erl -eval
> "io:format(hello, []), halt()." ; done > collected.txt 2>&1
> grep real collected.txt | sed -e 's/^.*m//g' -e s/s$//g > data.txt
>
> To collect the necessary data. On a FreeBSD system, you might need the
> jot(1) command rather than the seq(1) command, but the idea is the same.
>
> In R:
>
> x <- read.csv('data.txt', header=FALSE)
> > summary(x)
>        V1
>  Min.   :0.1890
>  1st Qu.:0.1930
>  Median :0.1940
>  Mean   :0.1948
>  3rd Qu.:0.1963
>  Max.   :0.2050
>
> In other words, my hunch was soundly destroyed by data. I've seen what I
> perceive to be longer boot times as well, and they seem to deviate more.
> But the human perception is so bad at precise measurements. Consider using
> the above script-idea together with R to obtain some more data and better
> numbers. The mean is usually only a good statistic if you have other data.
>
> --
> J.
>



-- 
J.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://erlang.org/pipermail/erlang-bugs/attachments/20150413/2bf02e44/attachment.htm>