Selectable heap allocators
cover Selectable heap allocators
Hi,
here comes a patch series for an idea I have in mind for quite some
time: break up the nucleus heap layer so that it can support different allocation schemes. Motivated was this idea by the great TLSF allocator by Miguel Masmano et al.
I once played with TLSF in userspace as a malloc replacement for RT apps, and I wondered if we shouldn't make use of it as well in the nucleus. It's most beautiful property is its strict O(1) complexity in my eyes, but it seems to provide even more qualities: less fragmentation, less overhead, maybe even better performance.
I haven't done any thorough analysis on those advantages (I'm counting on Miguel and others to prove them ;) ), only stability tests over the last weeks and a simple comparisons of the overhead. With 8 user-space tasks, a few mutexes, 8 RTDM sockets, and likely some other stuff allocated I get e.g. these usage numbers:
bsdalloc:
size=524288:used=17720:pagesz=512
tlsfalloc:
size=524288:used=14036:pagesz=512
This patch series now first tries to abstract a common interface for heap allocators. Those allocators can be statically selected at compile time. As the second step it adds latest TLSF version 2.2.0 as an alternative to BSD. Default remains BSD for now.
TLSF is currently limited to 32-bit archs, but 64-bit support is on the roadmap according to Miguel. He kindly helped me with preparing the TLSF patch for Xenomai usage. Thanks to him at this point!
Jan
Quoted 19 lines · Jan Kiszka
Hi,
here comes a patch series for an idea I have in mind for quite some
time: break up the nucleus heap layer so that it can support different allocation schemes. Motivated was this idea by the great TLSF allocator by Miguel Masmano et al.I once played with TLSF in userspace as a malloc replacement for RT apps, and I wondered if we shouldn't make use of it as well in the nucleus. It's most beautiful property is its strict O(1) complexity in my eyes, but it seems to provide even more qualities: less fragmentation, less overhead, maybe even better performance.
I haven't done any thorough analysis on those advantages (I'm counting on Miguel and others to prove them ;) ), only stability tests over the last weeks and a simple comparisons of the overhead. With 8 user-space tasks, a few mutexes, 8 RTDM sockets, and likely some other stuff allocated I get e.g. these usage numbers:
Quoted 6 lines
bsdalloc:
size=524288:used=17720:pagesz=512tlsfalloc:
size=524288:used=14036:pagesz=512
I'm ok to merge different allocators - and the needed infrastructure to support that - provided having several of them actually buys us something significant, maybe depending on some usage patterns. Actually, I once made the basic assumption that the BSDish heap could fulfill any kind of allocation requirements for all the use cases we should care about; maybe I was wrong, in which case, we should discuss of the particular issue of introducing an abstract layer to support several allocators. If the assumption remains unchallenged, then we don't need to abstract this, and the issue boils down to whether we should switch to TLSF or keep BSD.
This also means that we first need real figures for practical use cases comparing BSD and TLSF, before going any step further (at least internal and external fragmentation values, allocation and deallocation worst-case for bulks of contiguous and discontiguous blocks from various sizes, and anything else you may find relevant). At that point, we could decide whether we should provide both over the abstract infrastructure, simply replace BSD with TLSF, or keep the status quo.
Quoted 16 lines
This patch series now first tries to abstract a common interface for heap allocators. Those allocators can be statically selected at compile time. As the second step it adds latest TLSF version 2.2.0 as an alternative to BSD. Default remains BSD for now.
TLSF is currently limited to 32-bit archs, but 64-bit support is on the roadmap according to Miguel. He kindly helped me with preparing the TLSF patch for Xenomai usage. Thanks to him at this point!
Jan
_______________________________________________
Xenomai-core mailing list
[email protected]
https://mail.gna.org/listinfo/xenomai-core
Signature · 1 line
Philippe.
Quoted 38 lines · Philippe Gerum
Quoted 19 lines · Jan Kiszka
Hi,
here comes a patch series for an idea I have in mind for quite some
time: break up the nucleus heap layer so that it can support different allocation schemes. Motivated was this idea by the great TLSF allocator by Miguel Masmano et al.I once played with TLSF in userspace as a malloc replacement for RT apps, and I wondered if we shouldn't make use of it as well in the nucleus. It's most beautiful property is its strict O(1) complexity in my eyes, but it seems to provide even more qualities: less fragmentation, less overhead, maybe even better performance.
I haven't done any thorough analysis on those advantages (I'm counting on Miguel and others to prove them ;) ), only stability tests over the last weeks and a simple comparisons of the overhead. With 8 user-space tasks, a few mutexes, 8 RTDM sockets, and likely some other stuff allocated I get e.g. these usage numbers:
Quoted 6 lines
bsdalloc:
size=524288:used=17720:pagesz=512tlsfalloc:
size=524288:used=14036:pagesz=512I'm ok to merge different allocators - and the needed infrastructure to support that - provided having several of them actually buys us something significant, maybe depending on some usage patterns. Actually, I once made the basic assumption that the BSDish heap could fulfill any kind of allocation requirements for all the use cases we should care about; maybe I was wrong, in which case, we should discuss of the particular issue of introducing an abstract layer to support several allocators. If the assumption remains unchallenged, then we don't need to abstract this, and the issue boils down to whether we should switch to TLSF or keep BSD.
Quoted 7 lines
This also means that we first need real figures for practical use cases comparing BSD and TLSF, before going any step further (at least internal and external fragmentation values, allocation and deallocation worst-case for bulks of contiguous and discontiguous blocks from various sizes, and anything else you may find relevant). At that point, we could decide whether we should provide both over the abstract infrastructure, simply replace BSD with TLSF, or keep the status quo.
I agree that we definitely need numbers. And I could imagine that TLSF turns out to be the better choice in most, if not all categories.
But given the situation that a) TLSF is not yet a full BSD replacement due to the known shortcomings (missing 64 bit, no extensible heaps) while b) we want to test it against real Xenomai load, I think that having such switching infrastructure even just for one or two versions is the most effective way. Maintaining a patch just to allow testing would be a pain and would not attract further users beyond the core team.
Jan