Would smart variable ordering improve performance?
Hi! It is known that increasing data locality is a good way to benefit from CPU caches. There has been some effort from the IBMâs JVM crew that aligns some frequently accessed variables at a particular offset to improve accesses. Would this approach works in Pharo? Other similar situation: will the expression âObject new yourselfâ be faster if the object (i.e., result of Object new) is physically close to the class Object in memory? Cheers, Alexandre -- _,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;: Alexandre Bergel http://www.bergel.eu ^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;.
Hi Alexandre, On Fri, Oct 23, 2015 at 2:18 PM, Alexandre Bergel <alexandre.bergel@me.com> wrote:
Hi!
It is known that increasing data locality is a good way to benefit from CPU caches. There has been some effort from the IBMâs JVM crew that aligns some frequently accessed variables at a particular offset to improve accesses. Would this approach works in Pharo?
Other similar situation: will the expression âObject new yourselfâ be faster if the object (i.e., result of Object new) is physically close to the class Object in memory?
Clément and I have discussed this issue. It's one significant advantage that a conventional procedural language has over an OO one. Compare some rendering pipeline that in its OO version has to traverse deep object structure to dig out parameters as opposed to some array based implementation where the parameters are extracted out into separate arrays and hence the traversal is much simpler. I doubt that some minor rearrangement or alignment of structure would make any difference. But imagine an adaptive optimizer for data structure that could make the same transformation, automatically and invisibly to the programmer. Such a general purpose optimizer would be far more powerful. That's the kind of solution we think will be tried and that we would try if we encountered such performance issues and we had time to address it. But I think this is beyond the current state of the art and it feels like a PhD topic. I do remember Yoshiki saying that some plugin was doing something similar, but I could be mistaken.
Cheers, Alexandre -- _,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;: Alexandre Bergel http://www.bergel.eu ^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;.
-- _,,,^..^,,,_ best, Eliot
Glad to hear this. It would be fantastic if the VM would enable such experiment (e.g., relayouting variables and measuring impact on the low level cache). Enjoy Splash! Cheers, Alexandre
On Oct 26, 2015, at 1:29 AM, Eliot Miranda <eliot.miranda@gmail.com> wrote:
Hi Alexandre,
On Fri, Oct 23, 2015 at 2:18 PM, Alexandre Bergel <alexandre.bergel@me.com> wrote: Hi!
It is known that increasing data locality is a good way to benefit from CPU caches. There has been some effort from the IBMâs JVM crew that aligns some frequently accessed variables at a particular offset to improve accesses. Would this approach works in Pharo?
Other similar situation: will the expression âObject new yourselfâ be faster if the object (i.e., result of Object new) is physically close to the class Object in memory?
Clément and I have discussed this issue. It's one significant advantage that a conventional procedural language has over an OO one. Compare some rendering pipeline that in its OO version has to traverse deep object structure to dig out parameters as opposed to some array based implementation where the parameters are extracted out into separate arrays and hence the traversal is much simpler.
I doubt that some minor rearrangement or alignment of structure would make any difference. But imagine an adaptive optimizer for data structure that could make the same transformation, automatically and invisibly to the programmer. Such a general purpose optimizer would be far more powerful. That's the kind of solution we think will be tried and that we would try if we encountered such performance issues and we had time to address it.
But I think this is beyond the current state of the art and it feels like a PhD topic. I do remember Yoshiki saying that some plugin was doing something similar, but I could be mistaken.
Cheers, Alexandre -- _,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;: Alexandre Bergel http://www.bergel.eu ^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;.
-- _,,,^..^,,,_ best, Eliot
-- _,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;: Alexandre Bergel http://www.bergel.eu ^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;.
participants (2)
-
Alexandre Bergel -
Eliot Miranda