boost::flat_map and its performance compared to map and unordered_map
Solution 1
I have run a benchmark on different data structures very recently at my company so I feel I need to drop a word. It is very complicated to benchmark something correctly.
Benchmarking
On the web, we rarely find (if ever) a well-engineered benchmark. Until today I only found benchmarks that were done the journalist way (pretty quickly and sweeping dozens of variables under the carpet).
1) You need to consider cache warming
Most people running benchmarks are afraid of timer discrepancy, therefore they run their stuff thousands of times and take the whole time, they just are careful to take the same thousand of times for every operation, and then consider that comparable.
The truth is, in the real world it makes little sense, because your cache will not be warm, and your operation will likely be called just once. Therefore you need to benchmark using RDTSC, and time stuff calling them once only. Intel has made a paper describing how to use RDTSC (using a cpuid instruction to flush the pipeline, and calling it at least 3 times at the beginning of the program to stabilize it).
2) RDTSC accuracy measure
I also recommend doing this:
u64 g_correctionFactor; // number of clocks to offset after each measurement to remove the overhead of the measurer itself.
u64 g_accuracy;
static u64 const errormeasure = ~((u64)0);
#ifdef _MSC_VER
#pragma intrinsic(__rdtsc)
inline u64 GetRDTSC()
{
int a[4];
__cpuid(a, 0x80000000); // flush OOO instruction pipeline
return __rdtsc();
}
inline void WarmupRDTSC()
{
int a[4];
__cpuid(a, 0x80000000); // warmup cpuid.
__cpuid(a, 0x80000000);
__cpuid(a, 0x80000000);
// measure the measurer overhead with the measurer (crazy he..)
u64 minDiff = LLONG_MAX;
u64 maxDiff = 0; // this is going to help calculate our PRECISION ERROR MARGIN
for (int i = 0; i < 80; ++i)
{
u64 tick1 = GetRDTSC();
u64 tick2 = GetRDTSC();
minDiff = std::min(minDiff, tick2 - tick1); // make many takes, take the smallest that ever come.
maxDiff = std::max(maxDiff, tick2 - tick1);
}
g_correctionFactor = minDiff;
printf("Correction factor %llu clocks\n", g_correctionFactor);
g_accuracy = maxDiff - minDiff;
printf("Measurement Accuracy (in clocks) : %llu\n", g_accuracy);
}
#endif
This is a discrepancy measurer, and it will take the minimum of all measured values, to avoid getting a -10**18 (64 bits first negatives values) from time to time.
Notice the use of intrinsics and not inline assembly. First inline assembly is rarely supported by compilers nowadays, but much worse of all, the compiler creates a full ordering barrier around inline assembly because it cannot static analyze the inside, so this is a problem to benchmark real-world stuff, especially when calling stuff just once. So an intrinsic is suited here because it doesn't break the compiler free-re-ordering of instructions.
3) parameters
The last problem is people usually test for too few variations of the scenario. A container performance is affected by:
- Allocator
- size of the contained type
- cost of implementation of the copy operation, assignment operation, move operation, construction operation, of the contained type.
- number of elements in the container (size of the problem)
- type has trivial 3.-operations
- type is POD
Point 1 is important because containers do allocate from time to time, and it matters a lot if they allocate using the CRT "new" or some user-defined operation, like pool allocation or freelist or other...
(for people interested about pt 1, join the mystery thread on gamedev about system allocator performance impact)
Point 2 is because some containers (say A) will lose time copying stuff around, and the bigger the type the bigger the overhead. The problem is that when comparing to another container B, A may win over B for small types, and lose for larger types.
Point 3 is the same as point 2, except it multiplies the cost by some weighting factor.
Point 4 is a question of big O mixed with cache issues. Some bad-complexity containers can largely outperform low-complexity containers for a small number of types (like map
vs. vector
, because their cache locality is good, but map
fragments the memory). And then at some crossing point, they will lose, because the contained overall size starts to "leak" to main memory and cause cache misses, that plus the fact that the asymptotic complexity can start to be felt.
Point 5 is about compilers being able to elide stuff that are empty or trivial at compile time. This can optimize greatly some operations because the containers are templated, therefore each type will have its own performance profile.
Point 6 same as point 5, PODs can benefit from the fact that copy construction is just a memcpy
, and some containers can have a specific implementation for these cases, using partial template specializations, or SFINAE to select algorithms according to traits of T.
About the flat map
Apparently, the flat map is a sorted vector wrapper, like Loki AssocVector, but with some supplementary modernizations coming with C++11, exploiting move semantics to accelerate insert and delete of single elements.
This is still an ordered container. Most people usually don't need the ordering part, therefore the existence of unordered..
.
Have you considered that maybe you need a flat_unorderedmap
? which would be something like google::sparse_map
or something like that—an open address hash map.
The problem of open address hash maps is that at the time of rehash
they have to copy everything around to the new extended flat land, whereas a standard unordered map just has to recreate the hash index, while the allocated data stays where it is. The disadvantage of course is that the memory is fragmented like hell.
The criterion of a rehash in an open address hash map is when the capacity exceeds the size of the bucket vector multiplied by the load factor.
A typical load factor is 0.8
; therefore, you need to care about that, if you can pre-size your hash map before filling it, always pre-size to: intended_filling * (1/0.8) + epsilon
this will give you a guarantee of never having to spuriously rehash and recopy everything during filling.
The advantage of closed address maps (std::unordered..
) is that you don't have to care about those parameters.
But the boost::flat_map
is an ordered vector; therefore, it will always have a log(N) asymptotic complexity, which is less good than the open address hash map (amortized constant time). You should consider that as well.
Benchmark results
This is a test involving different maps (with int
key and __int64
/somestruct
as value) and std::vector
.
tested types information:
typeid=__int64 . sizeof=8 . ispod=yes
typeid=struct MediumTypePod . sizeof=184 . ispod=yes
Insertion
EDIT:
My previous results included a bug: they actually tested ordered insertion, which exhibited a very fast behavior for the flat maps.
I left those results later down this page because they are interesting.
This is the correct test:
I have checked the implementation, there is no such thing as a deferred sort implemented in the flat maps here. Each insertion sorts on the fly, therefore this benchmark exhibits the asymptotic tendencies:
map: O(N * log(N))
hashmaps: O(N)
vector and flatmaps: O(N * N)
Warning: hereafter the 2 tests for std::map
and both flat_map
s are buggy and actually test ordered insertion (vs random insertion for other containers. yes it's confusing sorry):
We can see that ordered insertion, results in back pushing, and is extremely fast. However, from the non-charted results of my benchmark, I can also say that this is not near the absolute optimality for a back-insertion. At 10k elements, perfect back-insertion optimality is obtained on a pre-reserved vector. Which gives us 3Million cycles; we observe 4.8M here for the ordered insertion into the flat_map
(therefore 160% of the optimal).
Analysis: remember this is 'random insert' for the vector, so the massive 1 billion cycles come from having to shift half (in average) the data upward (one element by one element) at each insertion.
Random search of 3 elements (clocks renormalized to 1)
in size = 100
in size = 10000
Iteration
over size 100 (only MediumPod type)
over size 10000 (only MediumPod type)
Final grain of salt
In the end, I wanted to come back on "Benchmarking §3 Pt1" (the system allocator). In a recent experiment, I am doing around the performance of an open address hash map I developed, I measured a performance gap of more than 3000% between Windows 7 and Windows 8 on some std::unordered_map
use cases (discussed here).
This makes me want to warn the reader about the above results (they were made on Win7): your mileage may vary.
Solution 2
From the docs it seems this is analogous to Loki::AssocVector
which I'm a fairly heavy user of. Since it's based on a vector it has the characteristics of a vector, that is to say:
- Iterators gets invalidated whenever
size
grows beyondcapacity
. - When it grows beyond
capacity
it needs to reallocated and move objects over, ie insertion isn't guaranteed constant time except for the special case of inserting atend
whencapacity > size
- Lookup is faster than
std::map
due to cache locality, a binary search which has the same performance characteristics asstd::map
otherwise - Uses less memory because it isn't a linked binary tree
- It never shrinks unless you forcibly tell it to ( since that triggers reallocation )
The best use is when you know the number of elements in advance ( so you can reserve
upfront ), or when insertion / removal is rare but lookup is frequent. Iterator invalidation makes it a bit cumbersome in some use cases so they're not interchangeable in terms of program correctness.
naumcho
Updated on July 23, 2022Comments
-
naumcho almost 2 years
It is common knowledge in programming that memory locality improves performance a lot due to cache hits. I recently found out about
boost::flat_map
which is a vector based implementation of a map. It doesn't seem to be nearly as popular as your typicalmap
/unordered_map
so I haven't been able to find any performance comparisons. How does it compare and what are the best use cases for it?Thanks!
-
Billy ONeal almost 10 yearsDoes
QueryPerformanceCounter
(or associated performance tools on other platforms) take care of this kind of thing for you? -
v.oddou almost 10 years
QueryPerformanceCounter
is actually quite a crazy machinery. And you cannot be sure of what it will use internally, it can use RDTSC, in this case it will normalize the readings so that it appears to have a stable 1Ghz frequency. It may use HPET and the call might be a bit lengthy because of that. Everything is written here: msdn.microsoft.com/en-us/library/windows/desktop/… . I recommendQueryPerfCntr
when you release code to clients. A benchmark is ran once on a controlled environment, use RDTSC, disable intel speed step, set thread prirority to time_critical -
v.oddou almost 10 yearsAlso don't take it like
QPC
will shield you from any harm, because it won't. Some words from the maker of virtual dub: virtualdub.org/blog/pivot/entry.php?id=106 . and a question from a person who notices a 1.4ms delay to return fromQPC
, this would completely break the benchmark : stackoverflow.com/questions/1723629/… -
Billy ONeal almost 10 yearsYour vector insert results include a sort or similar, right? I'd expect vector::push_back to beat map::insert
-
v.oddou almost 10 years@BillyONeal: You're absolutely right. I also realized after that these insertion results are super weird. There has to be a bug in my benchmark. I leave it like that for the moment, but I'm definitely going to investigate the code right now. I have a release going on though, so my time is scarse... raah
-
v.oddou almost 10 yearsI have re-run the benchmark, re checkd the code. everything seems fine. there is something i don't understand. The back-insertion with reservation in the vector takes 700,000 cycles, without pre-reserve it takes 3 Millions, so still faster than the 4.7 Millions of the boost::flatmap. However, the random insertion persists in being crazy slow. I use vector::insert with an iterator that I take from begin, call std::advance on it, with a random number. I would need to step-IN the stuffs to see what's going on and explain that.
-
Billy ONeal almost 10 yearsoh, in that case it makes sense. Vector's constant amortized time guarantees only apply when inserting at the end. Inserting at random positions should average O(n) per insert because everything after the insertion point must be moved forward. So we would expect quadratic behavior in your benchmark which blows up quite fast, even for small N. The AssocVector style implementations are probably deferring sort until a lookup is required, for example, rather than sorting after every insert. Hard to say without seeing your benchmark.
-
v.oddou almost 10 yearsOMG, I never thought of deferring the sort, that's amazing. That makes a lot of sense. Then I wonder why the boost documentation says that insertion is linear...
-
v.oddou almost 10 years@BillyONeal: Ah, we inspected the code with a colleague, and found the culprit, my "random" insertion was ordered because I used a std::set to make sure the inserted keys were unique. This is a plain imbecility, but I fixed that with a random_shuffle, I'm rebuilding now and some new results will appear soon as an edit. So the test in its current state prooves that the "ordered insertion" is damn fast.
-
Maxim Egorushkin almost 10 yearsDid you turn off CPU frequency scaling and turbo boost while benchmarking though? In Windows if you set it to "performance" it locks into turbo boost till the CPU overheats.
-
v.oddou almost 10 years@MaximYegorushkin: I set to "high performance", I verified with Core Temp that the frequency is locked to 3.6Ghz, But the turbo boost makes it vary to 3.7 from time to time. However, I'm not completely sure that it is a problem to let the CPU change frequency, since I measure everything in clock cycles, whatever the frequency, it won't change the results :) right ?
-
v.oddou almost 10 yearsAh no, it will change the behavior versus memory latency, a lower frequency will appear to have better results as soon as we tests stuff that get out of the cache. alright well, I fixed my frequency to between 3.6 and 3.7 so it should be OK anyway.
-
NoSenseEtAl over 9 yearsfalse :) measurements above show map being faster than flat_map for find operations, I guess boost ppl need to fix the implementation, but in theory you are right.
-
isomorphismes over 8 years"Intel has a paper" ←and here it is
-
Surt over 8 years@v.oddou, vector is missing in the random search graph.
-
ahcox almost 8 yearsThe Intel paper you reference states, "Assumptions: In this paper, all the results shown were obtained by running tests on a platform whose BIOS was optimized by removing every factor that could cause indeterminism. All power optimization, Intel Hyper-Threading technology, frequency scaling and turbo mode functionalities were turned off."
-
skgbanga over 7 yearsHey, do you think you can make this benchmark suite available on github? I have a custom hash map implementation that I want to benchmark against these numbers. Thanks!
-
v.oddou over 7 yearsI would love to. But it is not my property :'( As a consolation maybe you can try putting your container in the bench bed of my open source version of the same thing that you did. Here: sourceforge.net/p/cgenericopenaddresshashmap/code/ci/master/… It's not the same richness but still better than nothing.
-
boycy over 7 yearsPerhaps I'm missing something obvious, but I don't understand why the random search is slower on the
flat_map
compared tostd::map
- is anyone able to explain this result? -
v.oddou over 7 yearsI would explain it as a specific overhead of the boost implementation of this time, and not an intrinsic character of the
flat_map
as a container. Because theAska::
version is faster than thestd::map
lookup. Proving that there is room for optimization. The expected performance is asymptotically the same, but maybe slightly better thanks to cache locality. With high size sets they should converge. -
chtz about 7 yearsFor inserting lots of elements into a
flat_set
(without any look-ups in between), one should better construct a new one using the appropriate constructor (followed by a merge or swap) -
Ruslan over 6 yearsYour abscissa labels don't make sense. How do you conclude from large number of clock cycles that
std::vector
's back insertion is faster? Shouldn't clock cycles be proportional to time? -
v.oddou over 6 years@Ruslan yes. that's because I haven't uploaded the chart of the particular test that allowed me to conclude that. But it had a 3 million clocks for the mediumSizedPod in the 10k test.
-
Don Hatch almost 5 yearsFor asymptotic tendency of inserting N items into a hashmap, you don't need to say "amortized N"; it's just N. (Inserting 1 item is amortized 1, though.)
-
linuxfreebird about 4 years@v.oddou Excellent job on presenting this work. I have a question. When you iterate over the std map, I assume its done in the ordering of the keys. Is it possible to iterate over the std map elements in their stored order, instead of searching through each branch of the binary tree to present them in key order, so that the iterator is unordered but much faster?
-
v.oddou about 4 years@linuxfreebird thanks! from memory the iteration is
begin(); ++; !=end()
(same as what a range for would generate). There might be a way to speed up iteration by using a very specially designed allocator. But with a default alloc, all nodes are all over the place in memory. There is no way to iterate other than following the pointers, I can think of. You see what I mean? -
linuxfreebird about 4 years@v.oddou Thanks for the reply. I request you help with stackoverflow.com/questions/60815144/… . I think the BST I am looking for is implemented as an array like structure that encodes the pointers as indices in the array. The closet thing I found that does this is github.com/mpaland/avl_array . I am doing numerics on large arrays that need to double in size when inserting so memory allocation calls is kept to a minimum. I could use some help.
-
gast128 almost 4 yearsGood point about the random inserting. I still have here the case that a flat_map is faster than a map but they are now much closer. It seems that a flat_map is good in iteration and low on allocation cost but when you have to move heavy objects around in the container it can become awful. Perhaps that's even a tip of Chandler's talk (i.e. put heavy objects outside associative containers).
-
欢乐的Xiaox over 2 yearsHave you tried insert with hint for std::map? According to en.cppreference.com/w/cpp/container/map/insert, insert with a good hint may improve the performance for ordered element.
-
Carlo Wood about 2 yearsI wonder why you didn't find my
StopWatch
class... github.com/CarloWood/cwds/blob/master/benchmark.h#L134 Granted, itsmeasure
member function returns something from a warmed cache (many repeats); but if you just start() and stop() you get exactly what you were interested in.