Boost logo

Boost :

Subject: Re: [boost] Proposal: Monotonic Containers - Comparison with boost::pool and boost::fast_pool
From: Christian Schladetsch (christian.schladetsch_at_[hidden])
Date: 2009-06-19 01:58:46


Updated tests (more of them, better formatting of results) for
https://svn.boost.org/svn/boost/sandbox/monotonic/libs/monotonic/test/compare_memory_pool.cppis
here
http://tinyurl.com/m83vll for GCC and here http://tinyurl.com/n9g8jv for
MSVC. I also added a column for std::allocator/mono.

On Fri, Jun 19, 2009 at 1:31 PM, Christian Schladetsch <
christian.schladetsch_at_[hidden]> wrote:

> Hello,
>
> I ran the following test http://tinyurl.com/l89llq
>
> This compares std::allocator, boost::fast_pool_allocator,
> boost::pool_allocator and monotonic::allocator.
>
> The only test in which monotonic allocation fails to perform faster than
> the other schemes is thrash_pool_sort_list_int against fast_pool. The reason
> for this is that fast_pool effectively stores a pre-generated array of
> correctly-aligned chunks, so allocation against this is a pointer increment.
>
>
> Fast pool can do this, because it stores different pools for
> different-sized blocks. Rather, the monotonic allocator uses a single buffer
> for all allocations, so it has to align each time. There is an optimisation
> I can introduce between a monotonic::allocator<T> and the underlying storage
> that can improve performance slightly for this case.
>
> The downside for fast_pool is that the user has to keep track of each
> different-sized chunk that can possiblly be used, and it doesn't
> differentiate on type. From the documentation:
>
> "Since the size of T is used to determine the type of the underlying Pool,
> each allocator for different types of the same size will share the same
> underlying pool. The tag class prevents pools from being shared between
> pool_allocator and fast_pool_allocator. For example, on a system where
> sizeof(int) == sizeof(void *), pool_allocator<int> and pool_allocator<void
> *> will both allocate/deallocate from/to the same pool."
>
> This is problematic for cases where a given allocator is rebound to a
> different type, as happens for the node-based containers. I am unsure what
> the best practise is here, or if boost::pool and boost::fast_pool simply
> leak in this case.
>
> You will also not that fast_pool is very slow for thrash_pool_sort, which
> sorts different-sized vectors. Here, the same mechanism that made it %10-%20
> faster than monotonic using lists makes it many times slower when using
> vectors.
>
>
> In the following tables, the first row is the maximum number of elements in
> the relevant container; all internal columns are in seconds.
>
> Results for MSVC /O2.
>
> thrash_pool
> count fast_p pool std mono local fp/mono pool/mono
> 10 0.015 0.019 0.02 0.005 0.005 300% 380%
> 210 0.035 0.037 0.038 0.01 0.009 350% 370%
> 410 0.051 0.055 0.04 0.013 0.013 392.3% 423.1%
> 610 0.069 0.072 0.046 0.017 0.017 405.9% 423.5%
> 810 0.088 0.093 0.048 0.021 0.022 419% 442.9%
> 1010 0.105 0.111 0.053 0.026 0.028 403.8%
> 426.9%
> 1210 0.125 0.138 0.06 0.031 0.03 403.2%
> 445.2%
> 1410 0.142 0.143 0.061 0.034 0.033 417.6%
> 420.6%
> 1610 0.164 0.162 0.069 0.038 0.038 431.6%
> 426.3%
> 1810 0.179 0.184 0.072 0.043 0.042 416.3%
> 427.9%
> thrash_pool_iter
> count fast_p pool std mono local fp/mono pool/mono
> 10 0.027 0.032 0.032 0.012 0.011 225% 266.7%
> 210 0.152 0.146 0.118 0.102 0.101 149% 143.1%
> 410 0.27 0.263 0.199 0.191 0.189 141.4% 137.7%
> 610 0.4 0.372 0.286 0.283 0.284 141.3% 131.4%
> 810 0.509 0.487 0.376 0.372 0.372 136.8% 130.9%
> 1010 0.64 0.598 0.474 0.474 0.464 135%
> 126.2%
> 1210 0.765 0.721 0.579 0.563 0.592 135.9%
> 128.1%
> 1410 0.895 0.841 0.632 0.662 0.652 135.2%
> 127%
> 1610 0.991 0.923 0.701 0.714 0.759 138.8%
> 129.3%
> 1810 1.185 1.096 0.793 0.91 0.842 130.2%
> 120.4%
> thrash_pool_sort
> count fast_p pool std mono local fp/mono pool/mono
> 10 0.01 0.002 0.002 0.001 0.001 1000% 200%
> 110 0.091 0.011 0.011 0.008 0.009 1138% 137.5%
> 210 0.233 0.023 0.02 0.017 0.016 1371% 135.3%
> 310 0.495 0.034 0.026 0.025 0.025 1980% 136%
> 410 1.347 0.044 0.036 0.035 0.034 3849% 125.7%
> 510 1.411 0.058 0.045 0.042 0.042 3360% 138.1%
> 610 2.928 0.067 0.055 0.052 0.053 5631% 128.8%
> 710 2.54 0.078 0.064 0.065 0.061 3908% 120%
> 810 6.175 0.089 0.073 0.07 0.07 8821% 127.1%
> 910 3.976 0.104 0.081 0.082 0.08 4849% 126.8%
> thrash_pool_sort_list_int
> count fast_p pool std mono local fp/mono pool/mono
> 10 0.006 0.011 0.011 0.003 0.004 200% 366.7%
> 210 0.057 0.087 0.088 0.06 0.059 95% 145%
> 410 0.102 0.214 0.169 0.135 0.125 75.56% 158.5%
> 610 0.156 0.367 0.254 0.199 0.196 78.39% 184.4%
> 810 0.211 0.588 0.346 0.26 0.261 81.15% 226.2%
> 1010 0.269 0.837 0.459 0.351 0.35 76.64%
> 238.5%
> 1210 0.347 1.16 0.547 0.441 0.431 78.68%
> 263%
> 1410 0.406 1.504 0.632 0.54 0.511 75.19%
> 278.5%
> 1610 0.481 1.882 0.728 0.583 0.587 82.5%
> 322.8%
> 1810 0.559 2.32 0.817 0.671 0.674 83.31%
> 345.8%
> thrash_pool_map_list_unaligned
> count fast_p pool std mono local fp/mono pool/mono
> 10 0.002 0.003 0.004 0.001 0.001 200% 300%
> 210 0.033 0.083 0.079 0.029 0.029 113.8% 286.2%
> 410 0.067 0.214 0.164 0.058 0.061 115.5% 369%
> 610 0.102 0.404 0.246 0.091 0.092 112.1% 444%
> 810 0.139 0.646 0.338 0.122 0.126 113.9% 529.5%
> 1010 0.174 0.936 0.415 0.155 0.16 112.3%
> 603.9%
> 1210 0.219 1.271 0.51 0.19 0.197 115.3%
> 668.9%
> 1410 0.253 1.675 0.594 0.23 0.237 110%
> 728.3%
> 1610 0.297 2.144 0.678 0.273 0.274 108.8%
> 785.3%
> 1810 0.339 2.692 0.874 0.402 0.313 84.33%
> 669.7%
>
>
> Results for GCC -O2
> thrash_pool
> count fast_p pool std mono local fp/mono pool/mono
> 10 0 0.01 0 0.01 0 0% 100%
> 210 0.01 0.01 0.01 0 0 inf% inf%
> 410 0.01 0 0.01 0.01 0 100% 0%
> 610 0.01 0.01 0.01 0 0 inf% inf%
> 810 0.01 0 0.01 0.01 0 100% 0%
> 1010 0 0.01 0 0 0.01 nan%
> inf%
> 1210 0.01 0.01 0.01 0 0 inf%
> inf%
> 1410 0.01 0 0.01 0 0.01 inf%
> nan%
> 1610 0 0.01 0.01 0 0 nan%
> inf%
> 1810 0 0.01 0 0 0.01 nan%
> inf%
> thrash_pool_iter
> count fast_p pool std mono local fp/mono pool/mono
> 10 0.01 0.01 0 0.01 0 100% 100%
> 210 0.05 0.06 0.02 0.02 0.02 250% 300%
> 410 0.12 0.1 0.04 0.03 0.03 400% 333.3%
> 610 0.16 0.15 0.05 0.05 0.04 320% 300%
> 810 0.2 0.21 0.06 0.06 0.07 333.3% 350%
> 1010 0.25 0.25 0.07 0.08 0.08 312.5%
> 312.5%
> 1210 0.29 0.31 0.08 0.08 0.09 362.5%
> 387.5%
> 1410 0.33 0.35 0.1 0.08 0.1 412.5%
> 437.5%
> 1610 0.38 0.37 0.11 0.12 0.12 316.7%
> 308.3%
> 1810 0.43 0.45 0.12 0.13 0.13 330.8%
> 346.2%
> thrash_pool_sort
> count fast_p pool std mono local fp/mono pool/mono
> 10 0.01 0.01 0 0 0 inf% inf%
> 110 0.06 0.01 0.01 0.01 0.01 600% 100%
> 210 0.19 0.02 0.02 0.02 0.01 950% 100%
> 310 0.47 0.02 0.02 0.02 0.02 2350% 100%
> 410 0.69 0.04 0.04 0.03 0.02 2300% 133.3%
> 510 1.99 0.05 0.04 0.03 0.04 6633% 166.7%
> 610 1.65 0.06 0.04 0.04 0.04 4125% 150%
> 710 4.73 0.07 0.05 0.04 0.05 1.182e+04% 175%
> 810 2.28 0.07 0.07 0.06 0.06 3800% 116.7%
> 910 4.29 0.09 0.05 0.08 0.06 5362% 112.5%
> thrash_pool_sort_list_int
> count fast_p pool std mono local fp/mono pool/mono
> 10 0 0.01 0 0 0 nan% inf%
> 210 0.03 0.05 0.04 0.04 0.03 75% 125%
> 410 0.07 0.15 0.1 0.08 0.06 87.5% 187.5%
> 610 0.1 0.26 0.12 0.11 0.12 90.91% 236.4%
> 810 0.13 0.39 0.19 0.18 0.14 72.22% 216.7%
> 1010 0.17 0.57 0.24 0.19 0.2 89.47%
> 300%
> 1210 0.24 0.74 0.29 0.23 0.24 104.3%
> 321.7%
> 1410 0.26 0.98 0.36 0.26 0.28 100%
> 376.9%
> 1610 0.27 1.21 0.37 0.3 0.33 90%
> 403.3%
> 1810 0.35 1.51 0.43 0.35 0.37 100%
> 431.4%
> thrash_pool_map_list_unaligned
> count fast_p pool std mono local fp/mono pool/mono
> 10 0 0 0 0 0 nan% nan%
> 210 0.02 0.05 0.03 0.02 0.02 100% 250%
> 410 0.02 0.11 0.06 0.04 0.04 50% 275%
> 610 0.06 0.19 0.11 0.07 0.07 85.71% 271.4%
> 810 0.06 0.33 0.13 0.09 0.08 66.67% 366.7%
> 1010 0.09 0.51 0.17 0.11 0.11 81.82%
> 463.6%
> 1210 0.12 0.65 0.21 0.13 0.14 92.31%
> 500%
> 1410 0.15 0.93 0.25 0.14 0.14 107.1%
> 664.3%
> 1610 0.17 1.11 0.27 0.19 0.15 89.47%
> 584.2%
> 1810 0.18 1.35 0.33 0.21 0.21 85.71%
> 642.9%
>
>
> Regards,
> Christian
>
>


Boost list run by bdawes at acm.org, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk