Changeset 9cb6514 for doc/theses


Ignore:
Timestamp:
Aug 29, 2021, 10:58:58 PM (3 years ago)
Author:
Peter A. Buhr <pabuhr@…>
Branches:
ADT, ast-experimental, enum, forall-pointer-decay, jacob/cs343-translation, master, pthread-emulation, qualifiedEnum
Children:
01f78e0, 7737c29
Parents:
0be463c
Message:

proofread chapter performance.tex

File:
1 edited

Legend:

Unmodified
Added
Removed
  • doc/theses/andrew_beach_MMath/performance.tex

    r0be463c r9cb6514  
    1111Tests were run in \CFA, C++, Java and Python.
    1212In addition there are two sets of tests for \CFA,
    13 one for termination and once with resumption.
     13one with termination and one with resumption.
    1414
    1515C++ is the most comparable language because both it and \CFA use the same
     
    2121but otherwise \Cpp should have a significant advantage.
    2222
    23 Java a popular language with similar termination semantics, but
     23Java, a popular language with similar termination semantics, but
    2424it is implemented in a very different environment, a virtual machine with
    2525garbage collection.
     
    3838seem to be notable only for having resumption.
    3939Instead, resumption is compared to its simulation in other programming
    40 languages: fixup functions that are explicity passed into a function.
     40languages: fixup functions that are explicitly passed into a function.
    4141
    4242All tests are run inside a main loop that repeatedly performs a test.
    4343This approach avoids start-up or tear-down time from
    4444affecting the timing results.
    45 The number of times the loop is run is configurable from the command line,
     45The number of times the loop is run is configurable from the command line;
    4646the number used in the timing runs is given with the results per test.
    47 Tests ran their main loop a million times.
    48 The Java tests runs the main loop 1000 times before
     47% Tests ran their main loop a million times.
     48The Java tests run the main loop 1000 times before
    4949beginning the actual test to ``warm-up" the JVM.
    5050% All other languages are precompiled or interpreted.
     
    7272% \code{C++}{catch(...)}).
    7373
    74 When collecting data each test is run eleven times. The top three and bottom
     74When collecting data, each test is run eleven times. The top three and bottom
    7575three results are discarded and the remaining five values are averaged.
    76 The test are run with the latest (still pre-release) \CFA compiler was used,
     76The test are run with the latest (still pre-release) \CFA compiler,
    7777using gcc-10 as a backend.
    7878g++-10 is used for \Cpp.
     
    113113}
    114114\end{cfa}
    115 Other test cases have additional code around the recursive call add
     115Other test cases have additional code around the recursive call adding
    116116something besides simple stack frames to the stack.
    117 Note that both termination and resumption will have to traverse over
     117Note that both termination and resumption have to traverse over
    118118the stack but only termination has to unwind it.
    119119\begin{itemize}[nosep]
     
    124124\item Empty:
    125125The repeating function is empty except for the necessary control code.
    126 As other traversal tests add to this, so it is the baseline for the group
     126As other traversal tests add to this, it is the baseline for the group
    127127as the cost comes from traversing over and unwinding a stack frame
    128128that has no other interactions with the exception system.
     
    130130The repeating function creates an object with a destructor before calling
    131131itself.
    132 Comparing this to the empty test gives the time to traverse over and/or
     132Comparing this to the empty test gives the time to traverse over and
    133133unwind a destructor.
    134134\item Finally:
    135135The repeating function calls itself inside a try block with a finally clause
    136136attached.
    137 Comparing this to the empty test gives the time to traverse over and/or
     137Comparing this to the empty test gives the time to traverse over and
    138138unwind a finally clause.
    139139\item Other Handler:
    140140The repeating function calls itself inside a try block with a handler that
    141 will not match the raised exception, but is of the same kind of handler.
    142 This means that the EHM will have to check each handler, but will continue
    143 over all of the until it reaches the base of the stack.
    144 Comparing this to the empty test gives the time to traverse over and/or
     141does not match the raised exception, but is of the same kind of handler.
     142This means that the EHM has to check each handler, but continue
     143over all of them until it reaches the base of the stack.
     144Comparing this to the empty test gives the time to traverse over and
    145145unwind a handler.
    146146\end{itemize}
    147147
    148148\paragraph{Cross Try Statement}
    149 This group of tests measures the cost setting up exception handling if it is
     149This group of tests measures the cost for setting up exception handling, if it is
    150150not used (because the exceptional case did not occur).
    151 Tests repeatedly cross (enter and leave, execute) a try statement but never
    152 preform a raise.
     151Tests repeatedly cross (enter, execute, and leave) a try statement but never
     152perform a raise.
    153153\begin{itemize}[nosep]
    154154\item Handler:
     
    165165to handle that exception.
    166166
    167 There is the pattern shown in \CFA and \Cpp. Java and Python use the same
     167Here is the pattern shown in \CFA and \Cpp. Java and Python use the same
    168168pattern as \Cpp, but with their own syntax.
    169169
     
    229229going higher is better than going low) N, the number of iterations of the
    230230main loop in each test, is varied between tests. It is also given in the
    231 results and usually have a value in the millions.
     231results and has a value in the millions.
    232232
    233233An anomaly in some results came from \CFA's use of gcc nested functions.
    234234These nested functions are used to create closures that can access stack
    235235variables in their lexical scope.
    236 However, if they do so then they can cause the benchmark's run-time to
     236However, if they do so, then they can cause the benchmark's run-time to
    237237increase by an order of magnitude.
    238238The simplest solution is to make those values global variables instead
     
    309309% Now discuss the results in the tables.
    310310One result not directly related to \CFA but important to keep in mind is that,
    311 for exceptions the standard intuition about which languages should go
     311for exceptions, the standard intuition about which languages should go
    312312faster often does not hold.
    313313For example, there are a few cases where Python out-performs
     
    324324The only performance requirement is to insure the \CFA EHM has reasonable
    325325performance for prototyping.
    326 Although that may be hard to exactly quantify, we believe it has succeeded
     326Although that may be hard to exactly quantify, I believe it has succeeded
    327327in that regard.
    328328Details on the different test cases follow.
     329
     330\subsection{Termination, \autoref{t:PerformanceTermination}}
    329331
    330332\begin{description}
     
    332334\CFA is slower than \Cpp, but is still faster than the other languages
    333335and closer to \Cpp than other languages.
    334 This is to be expected as \CFA is closer to \Cpp than the other languages.
     336This result is to be expected as \CFA is closer to \Cpp than the other languages.
    335337
    336338\item[D'tor Traversal]
    337 Running destructors causes huge slowdown in every language that supports
     339Running destructors causes a huge slowdown in the two languages that support
    338340them. \CFA has a higher proportionate slowdown but it is similar to \Cpp's.
    339 Considering the amount of work done in destructors is so low the cost
    340 likely comes from the change of context required to do that work.
     341Considering the amount of work done in destructors is virtually zero (asm instruction), the cost
     342must come from the change of context required to trigger the destructor.
    341343
    342344\item[Finally Traversal]
    343 Speed is similar to Empty Traversal in all languages that support finally
     345Performance is similar to Empty Traversal in all languages that support finally
    344346clauses. Only Python seems to have a larger than random noise change in
    345347its run-time and it is still not large.
    346348Despite the similarity between finally clauses and destructors,
    347 finally clauses seem to avoid the spike in run-time destructors have.
     349finally clauses seem to avoid the spike that run-time destructors have.
    348350Possibly some optimization removes the cost of changing contexts.
    349351\todo{OK, I think the finally clause may have been optimized out.}
     
    354356This results in a significant jump.
    355357
    356 Other languages experiance a small increase in run-time.
     358Other languages experience a small increase in run-time.
    357359The small increase likely comes from running the checks,
    358360but they could avoid the spike by not having the same kind of overhead for
    359361switching to the check's context.
    360362
    361 \todo{Could revist Other Traversal, after Finally Traversal.}
     363\todo{Could revisit Other Traversal, after Finally Traversal.}
    362364
    363365\item[Cross Handler]
    364366Here \CFA falls behind \Cpp by a much more significant margin.
    365367This is likely due to the fact \CFA has to insert two extra function
    366 calls while \Cpp doesn't have to do execute any other instructions.
     368calls, while \Cpp does not have to execute any other instructions.
    367369Python is much further behind.
    368370
     
    370372\CFA's performance now matches \Cpp's from Cross Handler.
    371373If the code from the finally clause is being inlined,
    372 which is just a asm comment, than there are no additional instructions
     374which is just an asm comment, than there are no additional instructions
    373375to execute again when exiting the try statement normally.
    374376
    375377\item[Conditional Match]
    376 Both of the conditional matching tests can be considered on their own,
    377 however for evaluating the value of conditional matching itself the
     378Both of the conditional matching tests can be considered on their own.
     379However for evaluating the value of conditional matching itself, the
    378380comparison of the two sets of results is useful.
    379381Consider the massive jump in run-time for \Cpp going from match all to match
     
    384386possibly through resource reuse or their program representation.
    385387However \CFA is built like \Cpp and avoids the problem as well, this matches
    386 the pattern of the conditional match which makes the two execution paths
    387 much more similar.
     388the pattern of the conditional match, which makes the two execution paths
     389very similar.
    388390
    389391\end{description}
    390392
    391 Moving on to resumption there is one general note,
    392 resumption is \textit{fast}, the only test where it fell
     393\subsection{Resumption, \autoref{t:PerformanceResumption}}
     394
     395Moving on to resumption, there is one general note,
     396resumption is \textit{fast}. The only test where it fell
    393397behind termination is Cross Handler.
    394398In every other case, the number of iterations had to be increased by a
    395 factor of 10 to get the run-time in an approprate range
     399factor of 10 to get the run-time in an appropriate range
    396400and in some cases resumption still took less time.
    397401
     
    401405\item[Empty Traversal]
    402406See above for the general speed-up notes.
    403 This result is not surprising as resumption's link list approach
     407This result is not surprising as resumption's linked-list approach
    404408means that traversing over stack frames without a resumption handler is
    405409$O(1)$.
     
    408412Resumption does have the same spike in run-time that termination has.
    409413The run-time is actually very similar to Finally Traversal.
    410 As resumption does not unwind the stack both destructors and finally
    411 clauses are run while walking down the stack normally.
     414As resumption does not unwind the stack, both destructors and finally
     415clauses are run while walking down the stack during the recursion returns.
    412416So it follows their performance is similar.
    413417
    414418\item[Finally Traversal]
    415 The increase in run-time fromm Empty Traversal (once adjusted for
    416 the number of iterations) roughly the same as for termination.
    417 This suggests that the
     419% The increase in run-time from Empty Traversal (once adjusted for
     420% the number of iterations) is roughly the same as for termination.
     421% This suggests that the
     422See D'tor Traversal discussion.
    418423
    419424\item[Other Traversal]
     
    426431The only test case where resumption could not keep up with termination,
    427432although the difference is not as significant as many other cases.
    428 It is simply a matter of where the costs come from. Even if \CFA termination
    429 is not ``zero-cost" passing through an empty function still seems to be
    430 cheaper than updating global values.
     433It is simply a matter of where the costs come from. \PAB{What does this mean?
     434Even if \CFA termination
     435is not ``zero-cost", passing through an empty function still seems to be
     436cheaper than updating global values.}
    431437
    432438\item[Conditional Match]
     
    437443\end{description}
    438444
     445\subsection{Resumption/Fixup, \autoref{t:PerformanceFixupRoutines}}
     446
    439447Finally are the results of the resumption/fixup routine comparison.
    440 These results are surprisingly varied, it is possible that creating a closure
     448These results are surprisingly varied. It is possible that creating a closure
    441449has more to do with performance than passing the argument through layers of
    442450calls.
    443451Even with 100 stack frames though, resumption is only about as fast as
    444452manually passing a fixup routine.
     453However, as the number of fixup routines is increased, the cost of passing them
     454should make the resumption dynamic-search cheaper.
    445455So there is a cost for the additional power and flexibility exceptions
    446456provide.
Note: See TracChangeset for help on using the changeset viewer.