2.6.6. Weigh¶
Weigh is a tiny Haskell package
to measure allocations of data constructors and functions. It provides a similar
interface to Criterion, Gauge, and Tasty-Bench and is
useful to confirm that a data structure or function has the memory performance
you anticipate it to have at runtime. Weigh
is easy to setup and
non-invasive; requiring no changes to source code. Thus, it is a good initial
tool to use before trying more advanced methods with higher setup costs, such as
Cachegrind.
2.6.6.1. Requirements¶
The program must not be compiled with
-threaded
.The program must not be compiled with profiling enabled.
Weigh works by tracking the allocation and garbage collection behavior of the runtime system. It takes snapshots before and after forcing whatever value or function you are passing to it. Thus, it will report incorrect measurements if another thread changes the heap unexpectedly. Similarly, it will report larger results if the values are artificially inflated for profiling.
2.6.6.2. Restrictions¶
Weigh is portable, and should work anywhere GHC’s native runtime system will work.
2.6.6.3. What Information Do I Receive From Weigh?¶
Weigh reports a table similar to criterion
or gauge
except displays
allocations in bytes and garbage collections by default. For example:
Case Allocated GCs
() 0 0
1 0 0
True 0 0
[0..3] 192 0
Notice that ()
, 1
, True
do not allocate any memory. This is because
weigh
measures heap allocations. Thus, anything that GHC concludes can be
allocated at compile time, floated out of main
and treated like a constant
will return 0 allocations. The list allocates due to the use of List Ranges.
Contrast that with a static literal version:
Base Allocated GCs
() 0 0
1 0 0
True 0 0
[0..3] 192 0
[0,1,2,3] 0 0
which does no allocation as expected.
2.6.6.4. When should I use Weigh¶
Weigh is useful in the following scenarios:
Inspecting the memory footprint of a value, such as a data type. This could be useful to determine whether your data type will fit in a CPU cache line or not.
Inspecting the memory footprint of a function. This information could inform the decision to manually apply certain GHC optimizations that might not be firing. Such as Lambda Lifting or the SAT transformation.
Inspecting the memory footprint of a data structure or a function under a particular load. This is useful information to tune the data structure specifically for that load. For example, one might make tradeoffs that create slower writes in exchange for faster reads if the load requires exponentially more reads than writes.
Tracking the memory requirements and garbage collection pressure for phases or sub-systems of your program, or even defining unit tests based on the output of weigh.
Inspecting the performance differences of the same code compiled with different versions of GHC.
2.6.6.5. How should I use Weigh¶
Weigh’s interface is similar to criterion’s and gauge’s. Our recommendation is to define a specific benchmark target to keep your project tidy and modular. For example, for this chapter our toy programs are weighed through this cabal target:
benchmark weigh
type : exitcode-stdio-1.0
default-language: Haskell2010
ghc-options : -O2 -fforce-recomp
main-is : Main.hs
hs-source-dirs : bench/Weigh
build-depends: base >= 4.15
, containers
, deepseq
, weigh
, random
Then weigh
exports an API to run your code with. Here is a list of common
functions to use:
mainWith :: Weigh a -> IO ()
: Similar todefaultMain
in criterion. This function is the main entry point to run the tests.value :: NFData a => String -> a -> Weigh ()
: Measure the memory allocations of a single value.func :: NFData a => String -> (b -> a) -> b -> Weigh ()
: Measure the memory allocations that result from the input function.wgroup :: String -> Weigh () -> Weigh ()
: Define a group of tests.io :: NFData a => String -> (b -> IO a) -> b -> Weigh ()
: Weigh an IO action that is applied to the input argumentb
.io :: NFData a => String -> IO a -> Weigh ()
: Weigh an IO action.
We recommend using func
over value
because GHC might float out a given
value and statically allocate it. Thus the measurement will not observe the
allocation. For example, consider this program:
data Foo0 = Foo0
deriving (Generic,NFData)
data Foo1 = Foo1 Int
deriving (Generic,NFData)
data Foo2 = Foo2 String String
deriving (Generic,NFData)
one,two :: String
one = "one"
two = "two"
main :: IO ()
main = mainWith $ do
value "()" ()
value "1" (1 :: Int)
value "True" True
value "[0..3]" ([0..3] :: [Int])
value "[0,1,2,3]" ([0,1,2,3] :: [Int])
value "Foo0" Foo0
func "Foo1-func" Foo1 1
value "Foo1-value" (Foo1 1)
value "one" one
value "Foo2" (Foo2 one two)
One might expect ()
, 1
, and True
to be 0,
2 and 0 machine words respectively. However, this is not the case; here is the
output from weigh:
Running 1 benchmarks...
Benchmark weigh: RUNNING...
Case Allocated GCs
() 0 0
1 0 0
True 0 0
[0..3] 192 0
[0,1,2,3] 0 0
Foo0 0 0
Foo1-func 16 0
Foo1-value 0 0
one 144 0
Foo2 336 0
Benchmark weigh: FINISH
Notice that built in types such as ()
and True
do not do any allocation.
This is because these types are wired-in to
GHC, which means that there is a single shared ()
in GHC, so our call to
value "()" ()
performs no allocation because it references the shared
()
. This is also true for True
. 1
performs no allocation during the
runtime of our program because GHC realizes it is a CAF and float it
out of main
; and similarly so for [0,1,2,3]
and Foo0
. In contrast,
the list ranges [0..3]
and ([0,1,2],[3,4,5])
do perform allocation
during runtime. Foo1-func
allocates because we used func
which forces
the creation of Foo1
at runtime. Foo1-value
performs no runtime
allocation because GHC will detect it as a CAF just like True
, ()
and
1
. This is also why Foo2
allocates; under the hood the Foo2
value is
a CAF
, and its fields one
and two
are shared, but one
and
two
are allocated at runtime via the unpackCString#
primitive. Here is
the relevant Core:
-- RHS size: {terms: 3, types: 0, coercions: 0, joins: 0/0}
Main.main_v2 :: Foo2
[GblId,
Unf=Unf{Src=<vanilla>, TopLvl=True, Value=True, ConLike=True,
WorkFree=True, Expandable=True, Guidance=IF_ARGS [] 10 10}]
Main.main_v2 = Main.Foo2 one two
-- RHS size: {terms: 1, types: 0, coercions: 0, joins: 0/0}
Main.one1 :: ghc-prim:GHC.Prim.Addr#
[GblId,
Unf=Unf{Src=<vanilla>, TopLvl=True, Value=True, ConLike=True,
WorkFree=True, Expandable=True, Guidance=IF_ARGS [] 20 0}]
Main.one1 = "one"#
-- RHS size: {terms: 2, types: 0, coercions: 0, joins: 0/0}
one :: String
[GblId,
Unf=Unf{Src=<vanilla>, TopLvl=True, Value=False, ConLike=True,
WorkFree=False, Expandable=True, Guidance=IF_ARGS [] 20 0}]
one = ghc-prim:GHC.CString.unpackCString# Main.one1
Notice that Main.Foo2
calls one
and two
, and that one
is a CAF
because it has the properties Workfree=True
, Value=True
, and
TopLvl=True
. But the payload of one
, Main.one1
, is not a CAF
because it is not a value and is not work free.
2.6.6.6. Examples¶
2.6.6.6.1. Better Output By Setting Columns¶
Weigh’s default configuration is good enough, but we can query for more data
than just allocations and garbage collections using the function setColumns
:
main :: IO ()
main = mainWith $ do
setColumns [Case, Allocated, Max, Live, GCs, MaxOS] -- new
value "()" ()
value "1" (1 :: Int)
value "True" True
value "[0..3]" ([0..3] :: [Int])
value "[0,1,2,3]" ([0,1,2,3] :: [Int])
value "Foo0" Foo0
func "Foo1-func" Foo1 1
value "Foo1-value" (Foo1 1)
value "one" one
value "Foo2" (Foo2 one two)
which yields:
Running 1 benchmarks...
Benchmark weigh: RUNNING...
Case Allocated Max Live GCs MaxOS
() 0 456 456 0 0
1 0 456 456 0 0
True 0 456 456 0 0
[0..3] 192 504 504 0 0
[0,1,2,3] 0 456 456 0 0
Foo0 0 456 456 0 0
Foo1-func 16 472 472 0 0
Foo1-value 0 456 456 0 0
one 144 504 504 0 0
Foo2 336 552 552 0 0
Benchmark weigh: FINISH
and now we can see total bytes allocated, the maximum residency memory, the total amount of live data on the heap, the number of garbage collections, and the maximum memory in use by the RTS, which in these simple examples is always 0.
2.6.6.6.2. Observe GHC Optimizations on your Data Type¶
GHC is a good optimizing compiler and by using weigh we can investigate how GHC is optimizing our program. Consider these data types:
data LotsOfInts = A Int Int
| B Int Int
| C Int Int
| D Int Int
| E Int Int
deriving (Generic,NFData)
data LotsOfInts2 = A2 Int Int
| B2 Int Int
| C2 Int Int
| D2 Int Int
| E2 Int Int
| F2 Int Int
| G2 Int Int
| H2 Int Int
deriving (Generic,NFData)
With no optimizations taking place we would expect that a value of
LotsOfInts
requires 3 machine words: one for the data constructor header,
one for a pointer for each field. The payloads are simply Int
which requires
two machine words each, one for the I#
constructor and one for the payload
Int#
. Note that we could directly use unboxed Ints (that is store
Int#
instead of Int
in LotsOfInts
), each constructor would still
require three words only instead of storing pointers we would be storing the
actual payload. Let’s measure how many bytes a value of LotsOfInts
and
LotsOfInts2
requires:
{-# OPTIONS_GHC -O2 #-}
module Main where
import Weigh
data LotsOfInts = A Int Int
| B Int Int
| C Int Int
| D Int Int
| E Int Int
deriving (Generic,NFData)
data LotsOfInts2 = A2 Int Int
| B2 Int Int
| C2 Int Int
| D2 Int Int
| E2 Int Int
| F2 Int Int
| G2 Int Int
| H2 Int Int
| I2 Int Int
deriving (Generic,NFData)
-- notice that we'll use both value, and func to cover our bases
main :: IO ()
main = mainWith $ do
setColumns [Case, Allocated, Max, Live, GCs, MaxOS]
value "value: A Lot Of Ints" (A 1000 1001)
func "func: A Lot Of Ints" (A 1000) 1001
value "value: A Lot Of Ints2" (A2 1000 1001)
func "func: A Lot Of Ints2" (A2 1000) 1001
which produces:
Benchmark weigh: RUNNING...
Case Allocated Max Live GCs MaxOS
value: A Lot Of Ints 0 456 456 0 0
func: A Lot Of Ints 24 480 480 0 0
value: A Lot Of Ints2 72 456 456 0 0
func: A Lot Of Ints2 96 480 480 0 0
Benchmark weigh: FINISH
We expected LotsOfInts
to take a 3 machine words (or 24 bytes on a 64-bit
machine, 12 on a 32-bit machine) for a single value. However, we find that
instead of 24 bytes a new value allocates 0. In contrast, the function version
of LotsOfInts
allocates 24 bytes. There are several things happening here:
First, A 1000 1001
is recognized as a CAF by GHC and allocated at
compiled time; which is why we do not see any allocation measured by weigh at
runtime. Second, GHC allocates 24 bytes for the function version because
partially applying (A 1000)
to 1001
requires allocating 1001
(two
words) and a pointer to 1001
(one word).
We should expect that LotsOfInts
and LotsOfInts2
would have identical
memory characteristics because the only difference between them is the number of
constructors (five vs nine). We find that is not the case, instead
LotsOfInts2
allocates much more memory (3x!) than LotsOfInts
.
Furthermore, a value of LotsOfInts2
allocates 9 words (72 bytes), which is 6
more words than expected and indicates that the value was not detected as a
CAF. What we’re observing is the benefits of pointer-tagging optimizations (see
Marlow et al. [4]). When a datatype has 8 or less (4 or less on
32-bit machines) constructors GHC uses the last 3-bits (2-bits on 32-bit) of the
constructor’s pointer to store information about the constructor. This produces
a significant speedup and lowers memory costs in many ways: It allows the
runtime system to avoid checking the Info Table for arity and payload
information, which improves branch prediction and function calls. It speeds up
case-expressions because the runtime system does not need to enter the scrutinee
to determine the constructor. Thus, it is a good idea to keep the number of
constructors for crucial datatypes at 8 or less.
We can double check that pointer-tagging is the culprit by removing a
constructor from LotsOfInts2
:
data LotsOfInts2 = A2 Int Int
| B2 Int Int
| C2 Int Int
| D2 Int Int
| E2 Int Int
| F2 Int Int
| G2 Int Int
| H2 Int Int
-- now we have exactly 8 constructors
-- | I2 Int Int
deriving (Generic,NFData)
which produces:
Benchmark weigh: RUNNING...
Case Allocated Max Live GCs MaxOS
value: A Lot Of Ints 0 456 456 0 0
func: A Lot Of Ints 24 480 480 0 0
value: A Lot Of Ints2 0 456 456 0 0
func: A Lot Of Ints2 24 480 480 0 0
Benchmark weigh: FINISH
And now LotsOfInts2
behaves exactly like LotsOfInts
.
Note
If you check the STG for the nine constructor version of LotsOfInts2
you’ll find that the root of the allocations come from the Generic
and
NFData
instances. The functions these instances generate use an extra
case-expression to scrutinize a value of LotsOfInts2
, and in STG (and
Core) that extra case expression means extra allocation. In addition to the
extra case-expression, the nine constructor version disables inlining and SAT optimizations.
2.6.6.7. Summary¶
We have explored using the weigh
library to do a shallow inspection of the
memory performance of a data type or function. Weigh is easy to setup and low
weight, so it a useful tool before reaching for more invasive methods. One
should use weigh to collect data for memory regressions, to verify mental models
of GHC’s optimizations on data types in micro-benchmarks, or to inspect memory
behavior of a program under a particular load.
2.6.6.8. References and Further Reading¶
The FPComplete blog post on weigh