nedmalloc, again

[jal]

Mmmm, weirdness, recompiled today and it now runs. Doesn't give to much trust in nedmalloc :). "nedmalloc allocator is 1.402341 times faster than standard". Now time to plug in my own malloc...


JAL

[JamesM]

Bump... results plz?

[jal]

Yeah, sorry, no results yet, been busy with other stuff. Maybe next week. Anyone else has any results?


JAL

[Combuster]

I could plug PDClib's malloc, but knowing it is an inefficient alpha version it is guaranteed to lose big time. So I didn't bother.

[NickJohnson]

I'm working on my libc's malloc this week - I'll test it if I finish soon enough: it should work really well on lots of similar allocations. Where is the code for nedmalloc's benchmark suite?

[jal]

Where is the code for nedmalloc's benchmark suite?

In the nedmalloc zip-file. It's just one Google away.


JAL

[quanganht]

Something's wrong here. I cannot compile test.c on Ubuntu x64

[jal]

Something's wrong here. I cannot compile test.c on Ubuntu x64

What errors do you get?


JAL

[quanganht]

Something's wrong here. I cannot compile test.c on Ubuntu x64

What errors do you get?


JAL

Like this

Code: Select all

In file included from nedmalloc.c:63,
                 from test.c:8:
malloc.c.h: In function ‘mmap_resize’:
malloc.c.h:3704: warning: cast to pointer from integer of different size
malloc.c.h: In function ‘sys_trim’:
malloc.c.h:4137: warning: comparison between pointer and integer
/tmp/cc3lTAIT.o: In function `AllocCache':
test.c:(.text+0x8ad7): undefined reference to `pthread_setspecific'
/tmp/cc3lTAIT.o: In function `InitPool':
test.c:(.text+0x913a): undefined reference to `pthread_key_create'
test.c:(.text+0x9293): undefined reference to `pthread_key_delete'
/tmp/cc3lTAIT.o: In function `FindMSpace':
test.c:(.text+0x9acd): undefined reference to `pthread_setspecific'
/tmp/cc3lTAIT.o: In function `neddestroypool':
test.c:(.text+0x9d1b): undefined reference to `pthread_key_delete'
/tmp/cc3lTAIT.o: In function `neddisablethreadcache':
test.c:(.text+0x9f49): undefined reference to `pthread_getspecific'
test.c:(.text+0x9f6a): undefined reference to `pthread_setspecific'
test.c:(.text+0xa0d3): undefined reference to `pthread_setspecific'
/tmp/cc3lTAIT.o: In function `nedpmalloc':
test.c:(.text+0xa21d): undefined reference to `pthread_getspecific'
test.c:(.text+0xa29c): undefined reference to `pthread_setspecific'
/tmp/cc3lTAIT.o: In function `nedpcalloc':
test.c:(.text+0xa62d): undefined reference to `pthread_getspecific'
test.c:(.text+0xa6ac): undefined reference to `pthread_setspecific'
/tmp/cc3lTAIT.o: In function `nedprealloc':
test.c:(.text+0xaa52): undefined reference to `pthread_getspecific'
test.c:(.text+0xaad1): undefined reference to `pthread_setspecific'
/tmp/cc3lTAIT.o: In function `nedpfree':
test.c:(.text+0xacd3): undefined reference to `pthread_getspecific'
test.c:(.text+0xad52): undefined reference to `pthread_setspecific'
/tmp/cc3lTAIT.o: In function `nedpmemalign':
test.c:(.text+0xaedc): undefined reference to `pthread_getspecific'
test.c:(.text+0xaf5b): undefined reference to `pthread_setspecific'
/tmp/cc3lTAIT.o: In function `nedpindependent_calloc':
test.c:(.text+0xb604): undefined reference to `pthread_getspecific'
test.c:(.text+0xb683): undefined reference to `pthread_setspecific'
/tmp/cc3lTAIT.o: In function `nedpindependent_comalloc':
test.c:(.text+0xba70): undefined reference to `pthread_getspecific'
test.c:(.text+0xbaef): undefined reference to `pthread_setspecific'
/tmp/cc3lTAIT.o: In function `runtest':
test.c:(.text+0xc287): undefined reference to `pthread_create'
test.c:(.text+0xc304): undefined reference to `pthread_join'
test.c:(.text+0xc343): undefined reference to `pthread_create'
test.c:(.text+0xc388): undefined reference to `pthread_join'
collect2: ld returned 1 exit status

Sorry for slow responding, f**king Spamhaus blocked me

[Combuster]

That is not a compilation error.

Ever tried linking to the correct libraries? (read: -lpthread)

[quanganht]

Got it. 10 threads under Ubuntu 9.10 running x86-64 Core 2 Duo. This is what I have

Code: Select all

nedmalloc allocator is 0.508529 times faster than standard

WTF is it anyway ?!? Linux malloc out performed ned's ?

EDIT: With 5 threads:

Code: Select all

nedmalloc allocator is 0.711086 times faster than standard

EDIT2: I think I have gone mad to do this:

Code: Select all

20 threads:
nedmalloc allocator is 0.316950 times faster than standard

50 threads:
nedmalloc allocator is 0.171097 times faster than standard

100 threads:
nedmalloc allocator is 0.091263 times faster than standard

100 threads with -O3:
nedmalloc allocator is 0.076826 times faster than standard

The nedmalloc has a huge delay before it fire up, especially with more and more threads.

[jal]

(read: -lpthread)

Or just plain -pthread.


JAL

[xyzzy]

In testing against my allocator I found that the latest released version of nedmalloc performs pretty badly. After switching to nedmalloc SVN (revision 1130 - latest was broken for me), it started performing much better.

[NickJohnson]

I finally finished my allocator, and ported it to Linux. It's only single threaded, so all of these tests are with only 1 thread. It's all with "-O2 -fomit-frame-pointer -march=native" under gcc 4.3.4 and glibc 2.3.7. (glibc was compiled with "-O2 -fomit-frame-pointer -march=native" for fairness)

I changed the output format for the benchmark to just output the speed of each allocator, so I could compare them all at once. Here are the results:

Code: Select all

standard: 1960572 ops/sec
nedmalloc: 2784614 ops/sec
abmalloc: 2329786 ops/sec

Clearly, for single threads, nedmalloc > abmalloc (my malloc) > glibc's malloc, although it seems like nedmalloc scales badly from other results. If I set it to not allocate the larger blocks, and only the power-of-2 <512 byte chunks it says are typical of C++ allocations, the results are quite different:

Code: Select all

standard: 2245848 ops/sec
nedmalloc: 3269323 ops/sec
abmalloc: 4247631 ops/sec

I specifically designed my allocator for small allocations, so it isn't very surprising that it's twice as fast as glibc's malloc for them.

Does anyone else have results?