www.digitalmars.com         C & C++   DMDScript  

digitalmars.D - Speeding up compilation by using jemalloc

reply "weaselcat" <weaselcat gmail.com> writes:
Might be of use to someone, but I was looking for ways to speed 
up dmd's albeit already fast compilation times.

Just by dropping in jemalloc in place of glibc's malloc via 
LD_PRELOAD on my linux machine I saw a 10-15% drop in compilation 
times across the board. Not sure if this is common knowledge or 
not.

(Sorry if this takes away from your coffee time : ) )
Jan 22 2015
next sibling parent reply "Vladimir Panteleev" <vladimir thecybershadow.net> writes:
On Thursday, 22 January 2015 at 19:37:11 UTC, weaselcat wrote:

 Might be of use to someone, but I was looking for ways to speed 
 up dmd's albeit already fast compilation times.

 Just by dropping in jemalloc in place of glibc's malloc via 
 LD_PRELOAD on my linux machine I saw a 10-15% drop in 
 compilation times across the board. Not sure if this is common 
 knowledge or not.

 (Sorry if this takes away from your coffee time : ) )


I thought dmd used a bump-the-pointer approach, since it almost 
never frees memory?
Jan 22 2015
parent reply ketmar <ketmar ketmar.no-ip.org> writes:
On Thu, 22 Jan 2015 19:39:24 +0000, Vladimir Panteleev wrote:


 On Thursday, 22 January 2015 at 19:37:11 UTC, weaselcat wrote:

 Might be of use to someone, but I was looking for ways to speed up
 dmd's albeit already fast compilation times.

 Just by dropping in jemalloc in place of glibc's malloc via LD_PRELOAD
 on my linux machine I saw a 10-15% drop in compilation times across the
 board. Not sure if this is common knowledge or not.

 (Sorry if this takes away from your coffee time : ) )


=20
 I thought dmd used a bump-the-pointer approach, since it almost never
 frees memory?


it not frees, but it still does alot of `new`. seems that jmalloc is=20
better at it.=
Jan 22 2015
next sibling parent "H. S. Teoh via Digitalmars-d" <digitalmars-d puremagic.com> writes:
On Thu, Jan 22, 2015 at 07:59:23PM +0000, ketmar via Digitalmars-d wrote:

 On Thu, 22 Jan 2015 19:39:24 +0000, Vladimir Panteleev wrote:
 

 On Thursday, 22 January 2015 at 19:37:11 UTC, weaselcat wrote:

 Might be of use to someone, but I was looking for ways to speed up
 dmd's albeit already fast compilation times.

 Just by dropping in jemalloc in place of glibc's malloc via
 LD_PRELOAD on my linux machine I saw a 10-15% drop in compilation
 times across the board. Not sure if this is common knowledge or
 not.

 (Sorry if this takes away from your coffee time : ) )


 
 I thought dmd used a bump-the-pointer approach, since it almost
 never frees memory?


 
 it not frees, but it still does alot of `new`. seems that jmalloc is
 better at it.


Huh?! I thought dmd just uses a bump-the-pointer allocation?  How can
that possibly be slower that whatever jmalloc does?


T

-- 
Тише едешь, дальше будешь.
Jan 22 2015
prev sibling parent "Vladimir Panteleev" <vladimir thecybershadow.net> writes:
On Thursday, 22 January 2015 at 19:59:23 UTC, ketmar wrote:

 On Thu, 22 Jan 2015 19:39:24 +0000, Vladimir Panteleev wrote:


 On Thursday, 22 January 2015 at 19:37:11 UTC, weaselcat wrote:

 Might be of use to someone, but I was looking for ways to 
 speed up
 dmd's albeit already fast compilation times.

 Just by dropping in jemalloc in place of glibc's malloc via 
 LD_PRELOAD
 on my linux machine I saw a 10-15% drop in compilation times 
 across the
 board. Not sure if this is common knowledge or not.

 (Sorry if this takes away from your coffee time : ) )


 
 I thought dmd used a bump-the-pointer approach, since it 
 almost never
 frees memory?


 it not frees, but it still does alot of `new`. seems that 
 jmalloc is
 better at it.


https://github.com/D-Programming-Language/dmd/blob/master/src/root/rmem.c

Looks like it only does that for C++ `new`.
Jan 22 2015
prev sibling next sibling parent Andrei Alexandrescu <SeeWebsiteForEmail erdani.org> writes:
On 1/22/15 11:37 AM, weaselcat wrote:

 Might be of use to someone, but I was looking for ways to speed up dmd's
 albeit already fast compilation times.

 Just by dropping in jemalloc in place of glibc's malloc via LD_PRELOAD
 on my linux machine I saw a 10-15% drop in compilation times across the
 board. Not sure if this is common knowledge or not.

 (Sorry if this takes away from your coffee time : ) )


I think we should statically link jemalloc with dmd. Any takers? -- Andrei
Jan 22 2015
prev sibling next sibling parent "weaselcat" <weaselcat gmail.com> writes:
On Thursday, 22 January 2015 at 19:37:11 UTC, weaselcat wrote:

 Might be of use to someone, but I was looking for ways to speed 
 up dmd's albeit already fast compilation times.

 Just by dropping in jemalloc in place of glibc's malloc via 
 LD_PRELOAD on my linux machine I saw a 10-15% drop in 
 compilation times across the board. Not sure if this is common 
 knowledge or not.

 (Sorry if this takes away from your coffee time : ) )


I had some time to bench it on a project not mine

I used vibe.d BTW

avgtime -d -q -r10 dub build --compiler=dmd --force

------------------------
Total time (ms): 73167.1
Repetitions    : 10
Sample mode    : 7290 (3 ocurrences)
Median time    : 7300.78
Avg time       : 7316.71
Std dev.       : 32.5878
Minimum        : 7281.8
Maximum        : 7385.81
95% conf.int.  : [7252.84, 7380.59]  e = 63.871
99% conf.int.  : [7232.77, 7400.65]  e = 83.9407
EstimatedAvg95%: [7296.52, 7336.91]  e = 20.1978
EstimatedAvg99%: [7290.17, 7343.26]  e = 26.5444

LD_PRELOAD=/usr/lib/libjemalloc.so avgtime -d -q -r10 dub build 
--compiler=dmd --force

------------------------
Total time (ms): 64026.4
Repetitions    : 10
Sample mode    : 6390 (3 ocurrences)
Median time    : 6399.68
Avg time       : 6402.64
Std dev.       : 18.3341
Minimum        : 6365.8
Maximum        : 6431.14
95% conf.int.  : [6366.7, 6438.57]  e = 35.9342
99% conf.int.  : [6355.41, 6449.86]  e = 47.2256
EstimatedAvg95%: [6391.27, 6414]  e = 11.3634
EstimatedAvg99%: [6387.7, 6417.57]  e = 14.934



p.s., avgtime from dub is a nice tool.
Jan 22 2015
prev sibling next sibling parent reply "Wyatt" <wyatt.epp gmail.com> writes:
On Thursday, 22 January 2015 at 19:37:11 UTC, weaselcat wrote:

 Might be of use to someone, but I was looking for ways to speed 
 up dmd's albeit already fast compilation times.

 Just by dropping in jemalloc in place of glibc's malloc via 
 LD_PRELOAD on my linux machine I saw a 10-15% drop in 
 compilation times across the board. Not sure if this is common 
 knowledge or not.


Might be interesting to try some other implementations, too?  
TCMalloc[0] and nedmalloc[1] at least have permissive licenses.  
(Hoard[2] and Lockless[3] also seem interesting, but are GPL or 
worse.)

-Wyatt

[0] Part of gperf-tools https://code.google.com/p/gperftools/
[1] http://www.nedprod.com/programs/portable/nedmalloc/index.html
[2] http://www.hoard.org/
[3] http://locklessinc.com/downloads/
Jan 23 2015
next sibling parent Andrei Alexandrescu <SeeWebsiteForEmail erdani.org> writes:
On 1/23/15 8:59 AM, Wyatt wrote:

 On Thursday, 22 January 2015 at 19:37:11 UTC, weaselcat wrote:

 Might be of use to someone, but I was looking for ways to speed up
 dmd's albeit already fast compilation times.

 Just by dropping in jemalloc in place of glibc's malloc via LD_PRELOAD
 on my linux machine I saw a 10-15% drop in compilation times across
 the board. Not sure if this is common knowledge or not.


 Might be interesting to try some other implementations, too? TCMalloc[0]
 and nedmalloc[1] at least have permissive licenses. (Hoard[2] and
 Lockless[3] also seem interesting, but are GPL or worse.)

 -Wyatt

 [0] Part of gperf-tools https://code.google.com/p/gperftools/
 [1] http://www.nedprod.com/programs/portable/nedmalloc/index.html
 [2] http://www.hoard.org/
 [3] http://locklessinc.com/downloads/


Definitely measurements should help. From the many measurements we've 
run at Facebook, jemalloc would be overall a good bet. Also I work with 
Jason Evans so if there are issues we want improved I might ask him :o). 
-- Andrei
Jan 23 2015
prev sibling parent "weaselcat" <weaselcat gmail.com> writes:
On Friday, 23 January 2015 at 16:59:41 UTC, Wyatt wrote:

 On Thursday, 22 January 2015 at 19:37:11 UTC, weaselcat wrote:

 Might be of use to someone, but I was looking for ways to 
 speed up dmd's albeit already fast compilation times.

 Just by dropping in jemalloc in place of glibc's malloc via 
 LD_PRELOAD on my linux machine I saw a 10-15% drop in 
 compilation times across the board. Not sure if this is common 
 knowledge or not.


 Might be interesting to try some other implementations, too?  
 TCMalloc[0] and nedmalloc[1] at least have permissive licenses.
  (Hoard[2] and Lockless[3] also seem interesting, but are GPL 
 or worse.)

 -Wyatt

 [0] Part of gperf-tools https://code.google.com/p/gperftools/
 [1] 
 http://www.nedprod.com/programs/portable/nedmalloc/index.html
 [2] http://www.hoard.org/
 [3] http://locklessinc.com/downloads/


Hi,
I actually did try tcmalloc and I can't remember if it performed 
better than malloc or not, but it performed overall worse than 
jemalloc. I'm assuming because it's(as the name impllies), 
optimized for heavily threaded scenarios which DMD is not.
I didn't try the others because they're not readily available in 
my package manager.
Jan 23 2015
prev sibling parent Andrei Alexandrescu <SeeWebsiteForEmail erdani.org> writes:
On 1/22/15 11:37 AM, weaselcat wrote:

 Might be of use to someone, but I was looking for ways to speed up dmd's
 albeit already fast compilation times.

 Just by dropping in jemalloc in place of glibc's malloc via LD_PRELOAD
 on my linux machine I saw a 10-15% drop in compilation times across the
 board. Not sure if this is common knowledge or not.

 (Sorry if this takes away from your coffee time : ) )


Who can take this? https://issues.dlang.org/show_bug.cgi?id=14054

Andrei
Jan 26 2015