Do multiple passes of the groups example#318
Do multiple passes of the groups example#318AnyOldName3 wants to merge 1 commit intovsg-dev:masterfrom
Conversation
This should reveal when allocators have degraded performance over time or when reusing previously-freed space.
|
Thumbs up on having a good test for this performance issue, I do wonder if this example is getting to test specific. Perhaps another example just focused on the mutlipass test would be better. |
|
I have merged this PR as branch of vsgExamples: AnyOldName3-multipass-groups When I run it against VSG master I see: $ vsggroups
using VsgVisitor
type : vsg::Group
numNodes : 5592405
numBytes : 212511376
average node size : 38
numNodesVisited : 55924050
construction time : 0.194289
traversal time : 0.125244
destruction time : 0.144473
total time : 0.464007
Nodes constructed per second : 2.87839e+07
Nodes visited per second : 4.46519e+08
Nodes destructed per second : 3.87089e+07
using VsgVisitor
type : vsg::Group
numNodes : 5592405
numBytes : 212511376
average node size : 38
numNodesVisited : 55924050
construction time : 4.65186
traversal time : 0.13395
destruction time : 0.146522
total time : 4.93233
Nodes constructed per second : 1.20219e+06
Nodes visited per second : 4.17499e+08
Nodes destructed per second : 3.81677e+07
using VsgVisitor
type : vsg::Group
numNodes : 5592405
numBytes : 212511376
average node size : 38
numNodesVisited : 55924050
construction time : 4.81207
traversal time : 0.12715
destruction time : 0.146463
total time : 5.08568
Nodes constructed per second : 1.16216e+06
Nodes visited per second : 4.39828e+08
Nodes destructed per second : 3.81829e+07
average construction time : 3.2194
average traversal time : 0.128781
average destruction time : 0.145819
average total time : 3.49401
Average Nodes constructed per second : 1.73709e+06
Average Nodes visited per second : 4.34255e+08
Average Nodes destructed per second : 3.83516e+07Which looks like a 24 X slowdown of construction between the first and second passes. Ouch. Is this similar to what you are seeing with VSG master without your performance fix? |
|
Yep, 20-25x is what I'm typically seeing. |
|
I have stripped down and copied this vsggroups version to a new vsgreallocations example to make it clearer what is going on leaving the original vsggorups in place. I'll now close this PR as it's now redundant, the new example was merged with PR #327. |
2 participants
This should reveal when allocators have degraded performance over time or when reusing previously-freed space.
If the groups example's main purpose is allocator benchmarking, then the new default value of three passes is sensible as it avoids the risk that reuse of freed space won't get benchmarked when the allocator is modified, but otherwise it may be sensible to change the default to a single pass.