oprofile shows quite a bit more overhead in 2.4's ACL processing vs 2.3's. I'm thinking of streamlining things a bit, along these lines. Any thoughts?
This acl.c compiles, but a fair amount of slapd code needs to be altered to use it. I guess if we commit and move forward on it it will have to be for 2.5.
Howard Chu wrote:
oprofile shows quite a bit more overhead in 2.4's ACL processing vs 2.3's. I'm thinking of streamlining things a bit, along these lines. Any thoughts?
This acl.c compiles, but a fair amount of slapd code needs to be altered to use it. I guess if we commit and move forward on it it will have to be for 2.5.
It's all in the ACLCHECK branch now. Passes "make test" for me, haven't done any profiling on it yet. I suspect there's not much difference unless overlays or other ACL plugins are in use.
Howard Chu writes:
oprofile shows quite a bit more overhead in 2.4's ACL processing vs 2.3's. I'm thinking of streamlining things a bit, along these lines. Any thoughts?
Only a few loose ones:
A brief look shows some deep indirection, which may be harder for compilers to optimize due to possible aliasing among intervening writes. Maybe you should help by unpacking some more common expressions into local variables.
Beyond that, seeing "ACL optimization" I think "bytecodes". acl.c has many if/switch statements just to find out what an acl expression is doing. A flat 'switch(*bytecode++)' implementation should help, though I haven't looked closely and I'm not about to try anytime soon.
Of course, that points to a slapd.conf 'access' language which maps more directly to the bytecodes, or compiler techniques like common subexpression elimination and lazy evaluation... a mini Haskell implementation in slapd, yay:-)
Hallvard B Furuseth wrote:
Howard Chu writes:
oprofile shows quite a bit more overhead in 2.4's ACL processing vs 2.3's. I'm thinking of streamlining things a bit, along these lines. Any thoughts?
Only a few loose ones:
A brief look shows some deep indirection, which may be harder for compilers to optimize due to possible aliasing among intervening writes. Maybe you should help by unpacking some more common expressions into local variables.
Profiling showed that the ACLCHECK branch didn't have any improvement over HEAD. Looks like it may not be worth the trouble.
Beyond that, seeing "ACL optimization" I think "bytecodes". acl.c has many if/switch statements just to find out what an acl expression is doing. A flat 'switch(*bytecode++)' implementation should help, though I haven't looked closely and I'm not about to try anytime soon.
Of course, that points to a slapd.conf 'access' language which maps more directly to the bytecodes, or compiler techniques like common subexpression elimination and lazy evaluation... a mini Haskell implementation in slapd, yay:-)
Yeesh...
Hallvard B Furuseth wrote:
Howard Chu writes:
oprofile shows quite a bit more overhead in 2.4's ACL processing vs 2.3's. I'm thinking of streamlining things a bit, along these lines. Any thoughts?
Only a few loose ones:
A brief look shows some deep indirection, which may be harder for compilers to optimize due to possible aliasing among intervening writes. Maybe you should help by unpacking some more common expressions into local variables.
Profiling showed that the ACLCHECK branch didn't have any improvement over HEAD. Looks like it may not be worth the trouble.
Are you going to drop this?
p.
masarati@aero.polimi.it wrote:
Hallvard B Furuseth wrote:
Howard Chu writes:
oprofile shows quite a bit more overhead in 2.4's ACL processing vs 2.3's. I'm thinking of streamlining things a bit, along these lines. Any thoughts?
Only a few loose ones:
A brief look shows some deep indirection, which may be harder for compilers to optimize due to possible aliasing among intervening writes. Maybe you should help by unpacking some more common expressions into local variables.
Profiling showed that the ACLCHECK branch didn't have any improvement over HEAD. Looks like it may not be worth the trouble.
Are you going to drop this?
Yes, doesn't seem to offer any real benefit. If you can think of some other reason to keep it, we can perhaps talk about how to make it viable.
masarati@aero.polimi.it wrote:
Are you going to drop this?
Yes, doesn't seem to offer any real benefit. If you can think of some other reason to keep it, we can perhaps talk about how to make it viable.
Not necessarily a request to revitalize it; having a compact, neutral interface for access control arguments would ease extending access control in the future, if needed.
p.
masarati@aero.polimi.it wrote:
masarati@aero.polimi.it wrote:
Are you going to drop this?
Yes, doesn't seem to offer any real benefit. If you can think of some other reason to keep it, we can perhaps talk about how to make it viable.
Not necessarily a request to revitalize it; having a compact, neutral interface for access control arguments would ease extending access control in the future, if needed.
True. I guess if there's no performance degradation (and it looks like not, the performance impact was zero) then we can go ahead and adopt it.
-
Hallvard B Furuseth -
Howard Chu -
masarati＠aero.polimi.it