5.7. Running a compiled program¶
To make an executable program, the GHC system compiles your code and then links it with a non-trivial runtime system (RTS), which handles storage management, thread scheduling, profiling, and so on.
The RTS has a lot of options to control its behaviour. For example, you can change the context-switch interval, the default size of the heap, and enable heap profiling. These options can be passed to the runtime system in a variety of different ways; the next section (Setting RTS options) describes the various methods, and the following sections describe the RTS options themselves.
5.7.1. Setting RTS options¶
There are four ways to set RTS options:
- on the command line between
+RTS ... -RTS, when running the program (Setting RTS options on the command line)
- at compile-time, using
-with-rtsopts=⟨opts⟩(Setting RTS options at compile time)
- with the environment variable
GHCRTS(Setting RTS options with the GHCRTS environment variable)
- by overriding “hooks” in the runtime system (“Hooks” to change RTS behaviour)
184.108.40.206. Setting RTS options on the command line¶
If you set the
appropriately when linking (see Options affecting linking), you can give RTS
options on the command line when running your program.
When your Haskell program starts up, the RTS extracts command-line
arguments bracketed between
-RTS as its own. For example:
$ ghc prog.hs -rtsopts [1 of 1] Compiling Main ( prog.hs, prog.o ) Linking prog ... $ ./prog -f +RTS -H32m -S -RTS -h foo bar
The RTS will snaffle
-H32m -S for itself, and the remaining
-f -h foo bar will be available to your program if/when it
-RTS option is required if the runtime-system options extend to
the end of the command line, as in this example:
% hls -ltr /usr/etc +RTS -A5m
If you absolutely positively want all the rest of the options in a
command line to go to the program (and not the RTS), use a
--. The difference is that
--RTS will not be passed to
the program, while
As always, for RTS options that take ⟨size⟩s: If the last character of ⟨size⟩ is a K or k, multiply by 1024; if an M or m, by 1024*1024; if a G or G, by 1024^3. (And any wraparound in the counters is your fault!)
+RTS -? RTS option will print out the RTS
options actually available in your program (which vary, depending on how
Since GHC is itself compiled by GHC, you can change RTS options in
the compiler using the normal
+RTS ... -RTS combination. For instance, to set
the maximum heap size for a compilation to 128M, you would add
+RTS -M128m -RTS to the command line.
220.127.116.11. Setting RTS options at compile time¶
GHC lets you change the default RTS options for a program at compile
time, using the
-with-rtsopts flag (Options affecting linking). A common
use for this is to give your program a default heap and/or stack size
that is greater than the default. For example, to set
18.104.22.168. Setting RTS options with the
GHCRTS environment variable¶
-rtsoptsflag is set to something other than
ignoreAllwhen linking, RTS options are also taken from the environment variable
GHCRTS. For example, to set the maximum heap size to 2G for all GHC-compiled programs (using an
GHCRTS='-M2G' export GHCRTS
RTS options taken from the
GHCRTSenvironment variable can be overridden by options given on the command line.
Setting something like
GHCRTS=-M2G in your environment is a
handy way to avoid Haskell programs growing beyond the real memory in
your machine, which is easy to do by accident and can cause the machine
to slow to a crawl until the OS decides to kill the process (and you
hope it kills the right one).
22.214.171.124. “Hooks” to change RTS behaviour¶
GHC lets you exercise rudimentary control over certain RTS settings for any given program, by compiling in a “hook” that is called by the run-time system. The RTS contains stub definitions for these hooks, but by writing your own version and linking it on the GHC command line, you can override the defaults.
Owing to the vagaries of DLL linking, these hooks don’t work under Windows when the program is built dynamically.
126.96.36.199.1. Runtime events¶
You can change the messages printed when the runtime system “blows up,” e.g., on stack overflow. The hooks for these are as follows:
OutOfHeapHook(unsigned long, unsigned long)¶
The heap-overflow message.
The stack-overflow message.
The message printed if
188.8.131.52.2. Event log output¶
Furthermore GHC lets you specify the way event log data (see
⟨flags⟩) is written through a custom
A sink of event-log data.
EventLogWriter. This is optional.
writeEventLog(void *eventlog, size_t eventlog_size)¶
Hands buffered event log data to your event log writer. Return true on success. Required for a custom
Note that this function may be called by multiple threads simultaneously.
Flush buffers (if any) of your custom
EventLogWriter. This can be
Note that this function may be called by multiple threads simultaneously.
Called when event logging is about to stop. This can be
To use an
EventLogWriter the RTS API provides the following functions:
Query whether the current runtime system supports the eventlog (e.g. whether the current executable was linked with
-eventlog) and, if it is supported, whether it is currently logging.
startEventLogging(const EventLogWriter *writer)¶
Start logging events to the given
EventLogWriter. Returns true on success or false if another writer has already been configured.
Tear down the active
EVENTLOG_NOT_SUPPORTED: The runtime system wasn’t compiled with eventlog support.
EventLogWriterhas not yet been configured.
EventLogWriterhas been configured and is running.
5.7.2. Miscellaneous RTS options¶
If yes (the default), the RTS installs signal handlers to catch things like
Ctrl-C. This option is primarily useful for when you are using the Haskell code as a DLL, and want to set your own signal handlers.
Note that even with
--install-signal-handlers=no, the RTS interval timer signal is still enabled. The timer signal is either SIGVTALRM or SIGALRM, depending on the RTS configuration and OS capabilities. To disable the timer signal, use the
-V0RTS option (see
If yes (the default), the RTS on Windows installs exception handlers to catch unhandled exceptions using the Windows exception handling mechanism. This option is primarily useful for when you are using the Haskell code as a DLL, and don’t want the RTS to ungracefully terminate your application on errors such as segfaults.
If yes (the default), the RTS on Windows will generate a core dump on any crash. These dumps can be inspected using debuggers such as WinDBG. The dumps record all code, registers and threading information at the time of the crash. Note that this implies
If yes (the default), the RTS on Windows will generate a stack trace on crashes if exception handling are enabled. In order to get more information in compiled executables, C code or DLLs symbols need to be available.
If given, uses
MADV_FREEon platforms where this results in more accurate resident memory usage of the program as shown in memory usage reporting tools (e.g. the
Using this is expected to make the program slightly slower.
On Linux, MADV_FREE is newer and faster because it can avoid zeroing pages if they are re-used by the process later (see
man 2 madvise), but for the trade-off that memory inspection tools like
topwill not immediately reflect the freeing in their display of resident memory (RSS column): Only under memory pressure will Linux actually remove the freed pages from the process and update its RSS statistics. Until then, the pages show up as
man 5 proc).
The delayed RSS update can confuse programmers debugging memory issues, production memory monitoring tools, and end users who may complain about undue memory usage shown in reporting tools, so with this flag it can be turned off.
On 64-bit machines, the runtime linker usually needs to map object code into the low 2Gb of the address space, due to the x86_64 small memory model where most symbol references are 32 bits. The problem is that this 2Gb of address space can fill up, especially if you’re loading a very large number of object files into GHCi.
This flag offers a workaround, albeit a slightly convoluted one. To be able to load an object file outside of the low 2Gb, the object code needs to be compiled with
-fPIC -fexternal-dynamic-refs. When the
+RTS -xpflag is passed, the linker will assume that all object files were compiled with
-fPIC -fexternal-dynamic-refsand load them anywhere in the address space. It’s up to you to arrange that the object files you load (including all packages) were compiled in the right way. If this is not the case for an object, the linker will probably fail with an error message when the problem is detected.
On some platforms where PIC is always the case, e.g. macOS and OpenBSD on x86_64, and macOS and Linux on aarch64 this flag is enabled by default. One repercussion of this is that referenced system libraries also need to be compiled with
-fPICif we need to load them in the runtime linker.
This option is for working around memory allocation problems only. Do not use unless GHCi fails with a message like “
failed to mmap() memory below 2Gb”. Consider recompiling the objects with
-fPIC -fexternal-dynamic-refsand using the
-xpflag instead. If you need to use this option to get GHCi working on your machine, please file a bug.
On 64-bit machines, the RTS needs to allocate memory in the low 2Gb of the address space. Support for this across different operating systems is patchy, and sometimes fails. This option is there to give the RTS a hint about where it should be able to allocate memory in the low 2Gb of the address space. For example,
+RTS -xm20000000 -RTSwould hint that the RTS should allocate starting at the 0.5Gb mark. The default is to use the OS’s built-in support for allocating memory in the low 2Gb if available (e.g.
MAP_32BITon Linux), or otherwise
This option relates to allocation limits; for more about this see GHC.Conc.enableAllocationLimit. When a thread hits its allocation limit, the RTS throws an exception to the thread, and the thread gets an additional quota of allocation before the exception is raised again, the idea being so that the thread can execute its exception handlers. The
-xqcontrols the size of this additional quota.
5.7.3. RTS options to control the garbage collector¶
There are several options to give you precise control over garbage collection. Hopefully, you won’t need any of these in normal operation, but there are several things that can be tweaked for maximum performance.
Default: on Since: 8.10.2 Reverse: –nonmoving-gc
Uses the generational copying garbage collector for all generations. This is the default.
Default: off Since: 8.10.1 Reverse: –copying-gc
Enable the concurrent mark-and-sweep garbage collector for old generation collectors. Typically GHC uses a stop-the-world copying garbage collector for all generations. This can cause long pauses in execution during major garbage collections.
--nonmoving-gcenables the use of a concurrent mark-and-sweep garbage collector for oldest generation collections. Under this collection strategy oldest-generation garbage collection can proceed concurrently with mutation.
--nonmoving-gccannot be used with
Default: off Since: a long time ago Reverse: none
Uses a mark-region garbage collection strategy for the oldest-generation heap. Note that this cannot be used in conjunction with heap profiling (
-hT) unless linked against the profiling runtime system with
Default: off Since: 8.10.1
An alias for
Set the allocation area size used by the garbage collector. The allocation area (actually generation 0 step 0) is fixed and is never resized (unless you use
-H [⟨size⟩], below).
Optimal settings depend on the actual machine, program, and other RTS options. Increasing the allocation area size means worse cache behaviour but fewer garbage collections and less promotion.
In general settings >= 4MB can reduce performance in some cases, in particular for single threaded operation. However in a parallel setting increasing the allocation area to
16MB, or even
64MBcan increase gc throughput significantly.
With only 1 generation (e.g.
-G ⟨generations⟩) the
-Aoption specifies the minimum allocation area, since the actual size of the allocation area will be resized according to the amount of data in the heap (see
-F ⟨factor⟩, below).
When heap profiling using a smaller allocation area can increase accuracy as more frequent major garbage collections also results in more frequent heap snapshots
Sets the limit on the total size of “large objects” (objects larger than about 3KB) that can be allocated before a GC is triggered. By default this limit is the same as the
Large objects are not allocated from the normal allocation area set by the
-Aflag, which is why there is a separate limit for these. Large objects tend to be much rarer than small objects, so most programs hit the
-Alimit before the
-ALlimit. However, the
-Alimit is per-capability, whereas the
-ALlimit is global, so as
-Ngets larger it becomes more likely that we hit the
-ALlimit first. To counteract this, it might be necessary to use a larger
-ALlimit when using a large
To see whether you’re making good use of all the memory reseverd for the allocation area (
-N), look at the output of
+RTS -Sand check whether the amount of memory allocated between GCs is equal to
-N. If not, there are two possible remedies: use
-nto set a nursery chunk size, or use
-ALto increase the limit for large objects.
Set the minimum size of the old generation. The old generation is collected whenever it grows to this size or the value of the
-F ⟨factor⟩option multiplied by the size of the live data at the previous major collection, whichever is larger.
Default: 4m with
-A16mor larger, otherwise 0.
-n4m] When set to a non-zero value, this option divides the allocation area (
-Avalue) into chunks of the specified size. During execution, when a processor exhausts its current chunk, it is given another chunk from the pool until the pool is exhausted, at which point a collection is triggered.
This option is only useful when running in parallel (
-N2or greater). It allows the processor cores to make better use of the available allocation area, even when cores are allocating at different rates. Without
-n, each core gets a fixed-size allocation area specified by the
-A, and the first core to exhaust its allocation area triggers a GC across all the cores. This can result in a collection happening when the allocation areas of some cores are only partially full, so the purpose of the
-nis to allow cores that are allocating faster to get more of the allocation area. This means less frequent GC, leading a lower GC overhead for the same heap size.
This is particularly useful in conjunction with larger
-Avalues, for example
-A64m -n4mis a useful combination on larger core counts (8+).
Use a compacting algorithm for collecting the oldest generation. By default, the oldest generation is collected using a copying algorithm; this option causes it to be compacted in-place instead. The compaction algorithm is slower than the copying algorithm, but the savings in memory use can be considerable.
For a given heap size (using the
-H [⟨size⟩]option), compaction can in fact reduce the GC cost by allowing fewer GCs to be performed. This is more likely when the ratio of live data to heap size is high, say greater than 30%.
Compaction doesn’t currently work when a single generation is requested using the
Automatically enable compacting collection when the live data exceeds ⟨n⟩% of the maximum heap size (see the
-M ⟨size⟩option). Note that the maximum heap size is unlimited by default, so this option has no effect unless the maximum heap size is set with
This option controls the amount of memory reserved for the older generations (and in the case of a two space collector the size of the allocation area) as a factor of the amount of live data. For example, if there was 2M of live data in the oldest generation when we last collected it, then by default we’ll wait until it grows to 4M before collecting it again.
The default seems to work well here. If you have plenty of memory, it is usually better to use
-H [⟨size⟩]) than to increase
-F ⟨factor⟩setting will be automatically reduced by the garbage collector when the maximum heap size (the
-M ⟨size⟩setting) is approaching.
The inverse rate at which unused memory is returned to the OS when it is no longer needed. After a large amount of allocation the RTS will start by retaining a lot of allocated blocks in case it will need them again shortly but then it will gradually release them based on the
-Fd ⟨factor⟩. On each subsequent major collection which is not caused by a heap overflow a little more memory will attempt to be returned until the amount retained is similar to the amount of live bytes.
Increasing this factor will make the rate memory is returned slower, decreasing it will make memory be returned more eagerly. Setting it to 0 will disable the memory return (which will emulate the behaviour in releases prior to 9.2).
Set the number of generations used by the garbage collector. The default of 2 seems to be good, but the garbage collector can support any number of generations. Anything larger than about 4 is probably not a good idea unless your program runs for a long time, because the oldest generation will hardly ever get collected.
Specifying 1 generation with
+RTS -G1gives you a simple 2-space collector, as you would expect. In a 2-space collector, the
-A ⟨size⟩option specifies the minimum allocation area size, since the allocation area will grow with the amount of live data in the heap. In a multi-generational collector the allocation area is a fixed size (unless you use the
Default: 0 Since: 6.12.1
Use parallel GC in generation ⟨gen⟩ and higher. Omitting ⟨gen⟩ turns off the parallel GC completely, reverting to sequential GC.
The default parallel GC settings are usually suitable for parallel programs (i.e. those using GHC.Conc.par, Strategies, or with multiple threads). However, it is sometimes beneficial to enable the parallel GC for a single-threaded sequential program too, especially if the program has a large amount of heap data and GC is a significant fraction of runtime. To use the parallel GC in a sequential program, enable the parallel runtime with a suitable
-N ⟨x⟩option, and additionally it might be beneficial to restrict parallel GC to the old generation with
Default: 1 for
-A< 32M, 0 otherwise
Use load-balancing in the parallel GC in generation ⟨gen⟩ and higher. Omitting ⟨gen⟩ disables load-balancing entirely.
Load-balancing shares out the work of GC between the available cores. This is a good idea when the heap is large and we need to parallelise the GC work, however it is also pessimal for the short young-generation collections in a parallel program, because it can harm locality by moving data from the cache of the CPU where is it being used to the cache of another CPU. Hence the default is to do load-balancing only in the old-generation. In fact, for a parallel program it is sometimes beneficial to disable load-balancing entirely with
Default: the value of
-Nor the number of CPU cores, whichever is smaller.
By default, all of the capabilities participate in parallel garbage collection. If we want to use a very large
-Nvalue, however, this can reduce the performance of the GC. For this reason, the
-qnflag can be used to specify a lower number for the threads that should participate in GC. During GC, if there are more than this number of workers active, some of them will sleep for the duration of the GC.
-qnflag may be useful when running with a large
-Avalue (so that GC is infrequent), and a large
-Nvalue (so as to make use of hyperthreaded cores, for example). For example, on a 24-core machine with 2 hyperthreads per core, we might use
-N48 -qn24 -A128mto specify that the mutator should use hyperthreads but the GC should only use real cores. Note that this configuration would use 6GB for the allocation area.
This option provides a “suggested heap size” for the garbage collector. Think of
-Hsizeas a variable
-A ⟨size⟩option. It says: I want to use at least ⟨size⟩ bytes, so use whatever is left over to increase the
This option does not put a limit on the heap size: the heap may grow beyond the given size as usual.
If ⟨size⟩ is omitted, then the garbage collector will take the size of the heap at the previous GC as the ⟨size⟩. This has the effect of allowing for a larger
-Avalue but without increasing the overall memory requirements of the program. It can be useful when the default small
-Avalue is suboptimal, as it can be in programs that create large amounts of long-lived data.
Default: 0.3 seconds in the threaded runtime, 0 in the non-threaded runtime
In the threaded and SMP versions of the RTS (see
-threaded, Options affecting linking), a major GC is automatically performed if the runtime has been idle (no Haskell computation has been running) for a period of time. The amount of idle time which must pass before a GC is performed is set by the
-I ⟨seconds⟩option. Specifying
-I0disables the idle GC.
For an interactive application, it is probably a good idea to use the idle GC, because this will allow finalizers to run and deadlocked threads to be detected in the idle time when no Haskell computation is happening. Also, it will mean that a GC is less likely to happen when the application is busy, and so responsiveness may be improved. However, if the amount of live data in the heap is particularly large, then the idle GC can cause a significant delay, and too small an interval could adversely affect interactive responsiveness.
The idle period timer only resets after some activity by a Haskell thread. If your program is doing literally nothing then after the first idle collection is triggered then no more future collections will be scheduled until more work is performed.
This is an experimental feature, please let us know if it causes problems and/or could benefit from further tuning.
Default: 0 seconds
By default, if idle GC is enabled in the threaded runtime, a major GC will be performed every time the process goes idle for a sufficiently long duration (see
-I ⟨seconds⟩). For large server processes accepting regular but infrequent requests (e.g., once per second), an expensive, major GC may run after every request. As an alternative to shutting off idle GC entirely (with
-I0), a minimum wait time between idle GCs can be specified with this flag. For example,
-Iw60will ensure that an idle GC runs at most once per minute.
This is an experimental feature, please let us know if it causes problems and/or could benefit from further tuning.
Set the initial stack size for new threads.
Thread stacks (including the main thread’s stack) live on the heap. As the stack grows, new stack chunks are added as required; if the stack shrinks again, these extra stack chunks are reclaimed by the garbage collector. The default initial stack size is deliberately small, in order to keep the time and space overhead for thread creation to a minimum, and to make it practical to spawn threads for even tiny pieces of work.
This flag used to be simply
-k, but was renamed to
-kiin GHC 7.2.1. The old name is still accepted for backwards compatibility, but that may be removed in a future version.
Set the size of “stack chunks”. When a thread’s current stack overflows, a new stack chunk is created and added to the thread’s stack, until the limit set by
-K ⟨size⟩is reached.
The advantage of smaller stack chunks is that the garbage collector can avoid traversing stack chunks if they are known to be unmodified since the last collection, so reducing the chunk size means that the garbage collector can identify more stack as unmodified, and the GC overhead might be reduced. On the other hand, making stack chunks too small adds some overhead as there will be more overflow/underflow between chunks. The default setting of 32k appears to be a reasonable compromise in most cases.
Sets the stack chunk buffer size. When a stack chunk overflows and a new stack chunk is created, some of the data from the previous stack chunk is moved into the new chunk, to avoid an immediate underflow and repeated overflow/underflow at the boundary. The amount of stack moved is set by the
Note that to avoid wasting space, this value should typically be less than 10% of the size of a stack chunk (
-kc ⟨size⟩), because in a chain of stack chunks, each chunk will have a gap of unused space of this size.
Default: 80% of physical memory
Set the maximum stack size for an individual thread to ⟨size⟩ bytes. If the thread attempts to exceed this limit, it will be sent the
StackOverflowexception. The limit can be disabled entirely by specifying a size of zero.
This option is there mainly to stop the program eating up all the available memory in the machine if it gets into an infinite loop.
Minimum % ⟨n⟩ of heap which must be available for allocation.
Set the maximum heap size to ⟨size⟩ bytes. The heap normally grows and shrinks according to the memory requirements of the program. The only reason for having this option is to stop the heap growing without bound and filling up all the available swap space, which at the least will result in the program being summarily killed by the operating system.
The maximum heap size also affects other garbage collection parameters: when the amount of live data in the heap exceeds a certain fraction of the maximum heap size, compacting collection will be automatically enabled for the oldest generation, and the
-Fparameter will be reduced in order to avoid exceeding the maximum heap size.
If the program’s heap exceeds the value set by
-M ⟨size⟩, the RTS throws an exception to the program, and the program gets an additional quota of allocation before the exception is raised again, the idea being so that the program can execute its exception handlers.
-Mgrace=controls the size of this additional quota.
Enable NUMA-aware memory allocation in the runtime (only available with
-threaded, and only on Linux and Windows currently).
Background: some systems have a Non-Uniform Memory Architecture, whereby main memory is split into banks which are “local” to specific CPU cores. Accessing local memory is faster than accessing remote memory. The OS provides APIs for allocating local memory and binding threads to particular CPU cores, so that we can ensure certain memory accesses are using local memory.
--numaoption tells the RTS to tune its memory usage to maximize local memory accesses. In particular, the RTS will:
- Determine the number of NUMA nodes (N) by querying the OS.
- Manage separate memory pools for each node.
- Map capabilities to NUMA nodes. Capability C is mapped to NUMA node C mod N.
- Bind worker threads on a capability to the appropriate node.
- Allocate the nursery from node-local memory.
- Perform other memory allocation, including in the GC, from node-local memory.
- When load-balancing, we prefer to migrate threads to another Capability on the same node.
--numaflag is typically beneficial when a program is using all cores of a large multi-core NUMA system, with a large allocation area (
-A). All memory accesses to the allocation area will go to local memory, which can save a significant amount of remote memory access. A runtime speedup on the order of 10% is typical, but can vary a lot depending on the hardware and the memory behaviour of the program.
Note that the RTS will not set CPU affinity for bound threads and threads entering Haskell from C/C++, so if your program uses bound threads you should ensure that each bound thread calls the RTS API rts_setInCallCapability(c,1) from C/C++ before calling into Haskell. Otherwise there could be a mismatch between the CPU that the thread is running on and the memory it is using while running Haskell code, which will negate any benefits of
If given an explicit <mask>, the <mask> is interpreted as a bitmap that indicates the NUMA nodes on which to run the program. For example,
--numa=3would run the program on NUMA nodes 0 and 1.
When a GC starts, all the running mutator threads have to stop and synchronise. The period between when the GC is initiated and all the mutator threads are stopped is called the GC synchronisation phase. If this phase is taking a long time (longer than 1ms is considered long), then it can have a severe impact on overall throughput.
A long GC sync can be caused by a mutator thread that is inside an
unsafeFFI call, or running in a loop that doesn’t allocate memory and so doesn’t yield. To fix the former, make the call
safe, and to fix the latter, either avoid calling the code in question or compile it with
By default, the flag will cause a warning to be emitted to stderr when the sync time exceeds the specified time. This behaviour can be overridden, however: the
longGCSync()hook is called when the sync time is exceeded during the sync period, and the
longGCSyncEnd()hook at the end. Both of these hooks can be overridden in the
RtsConfigwhen the runtime is started with
hs_init_ghc(). The default implementations of these hooks (
LongGCSyncEnd()respectively) print warnings to stderr.
One way to use this flag is to set a breakpoint on
LongGCSync()in the debugger, and find the thread that is delaying the sync. You probably want to use
-gto provide more info to the debugger.
The GC sync time, along with other GC stats, are available by calling the
getRTSStats()function from C, or
5.7.4. RTS options to produce runtime statistics¶
These options produce runtime-system statistics, such as the amount of time spent executing the program and in the garbage collector, the amount of memory allocated, the maximum size of the heap, and so on. The three variants give different levels of detail:
-Tcollects the data but produces no output
-tproduces a single line of output in the same format as GHC’s
-sproduces a more detailed summary at the end of the program, and
-Sadditionally produces information about each and every garbage collection. Passing
--internal-countersto a threaded runtime will cause a detailed summary to include various internal counts accumulated during the run; note that these are unspecified and may change between releases.
The output is placed in ⟨file⟩. If ⟨file⟩ is omitted, then the output is sent to
If you use the
-Tflag then, you should access the statistics using GHC.Stats.
If you use the
-tflag then, when your program finishes, you will see something like this:
<<ghc: 36169392 bytes, 69 GCs, 603392/1065272 avg/max bytes residency (2 samples), 3M in use, 0.00 INIT (0.00 elapsed), 0.02 MUT (0.02 elapsed), 0.07 GC (0.07 elapsed) :ghc>>
This tells you:
- The total number of bytes allocated by the program over the whole run.
- The total number of garbage collections performed.
- The average and maximum “residency”, which is the amount of live
data in bytes. The runtime can only determine the amount of live
data during a major GC, which is why the number of samples
corresponds to the number of major GCs (and is usually relatively
small). To get a better picture of the heap profile of your
program, use the
-hTRTS option (RTS options for profiling).
- The peak memory the RTS has allocated from the OS.
- The amount of CPU time and elapsed wall clock time while initialising the runtime system (INIT), running the program itself (MUT, the mutator), and garbage collecting (GC).
You can also get this in a more future-proof, machine readable format, with
[("bytes allocated", "36169392") ,("num_GCs", "69") ,("average_bytes_used", "603392") ,("max_bytes_used", "1065272") ,("num_byte_usage_samples", "2") ,("peak_megabytes_allocated", "3") ,("init_cpu_seconds", "0.00") ,("init_wall_seconds", "0.00") ,("mutator_cpu_seconds", "0.02") ,("mutator_wall_seconds", "0.02") ,("GC_cpu_seconds", "0.07") ,("GC_wall_seconds", "0.07") ]
If you use the
-sflag then, when your program finishes, you will see something like this (the exact details will vary depending on what sort of RTS you have, e.g. you will only see profiling data if your RTS is compiled for profiling):
36,169,392 bytes allocated in the heap 4,057,632 bytes copied during GC 1,065,272 bytes maximum residency (2 sample(s)) 54,312 bytes maximum slop 3 MB total memory in use (0 MB lost due to fragmentation) Generation 0: 67 collections, 0 parallel, 0.04s, 0.03s elapsed Generation 1: 2 collections, 0 parallel, 0.03s, 0.04s elapsed SPARKS: 359207 (557 converted, 149591 pruned) INIT time 0.00s ( 0.00s elapsed) MUT time 0.01s ( 0.02s elapsed) GC time 0.07s ( 0.07s elapsed) EXIT time 0.00s ( 0.00s elapsed) Total time 0.08s ( 0.09s elapsed) %GC time 89.5% (75.3% elapsed) Alloc rate 4,520,608,923 bytes per MUT second Productivity 10.5% of total user, 9.1% of total elapsed
The “bytes allocated in the heap” is the total bytes allocated by the program over the whole run.
GHC uses a copying garbage collector by default. “bytes copied during GC” tells you how many bytes it had to copy during garbage collection.
The maximum space actually used by your program is the “bytes maximum residency” figure. This is only checked during major garbage collections, so it is only an approximation; the number of samples tells you how many times it is checked.
The “bytes maximum slop” tells you the most space that is ever wasted due to the way GHC allocates memory in blocks. Slop is memory at the end of a block that was wasted. There’s no way to control this; we just like to see how much memory is being lost this way.
The “total memory in use” tells you the peak memory the RTS has allocated from the OS.
Next there is information about the garbage collections done. For each generation it says how many garbage collections were done, how many of those collections were done in parallel, the total CPU time used for garbage collecting that generation, and the total wall clock time elapsed while garbage collecting that generation.
SPARKSstatistic refers to the use of
Control.Parallel.parand related functionality in the program. Each spark represents a call to
par; a spark is “converted” when it is executed in parallel; and a spark is “pruned” when it is found to be already evaluated and is discarded from the pool by the garbage collector. Any remaining sparks are discarded at the end of execution, so “converted” plus “pruned” does not necessarily add up to the total.
Next there is the CPU time and wall clock time elapsed broken down by what the runtime system was doing at the time. INIT is the runtime system initialisation. MUT is the mutator time, i.e. the time spent actually running your code. GC is the time spent doing garbage collection. RP is the time spent doing retainer profiling. PROF is the time spent doing other profiling. EXIT is the runtime system shutdown time. And finally, Total is, of course, the total.
%GC time tells you what percentage GC is of Total. “Alloc rate” tells you the “bytes allocated in the heap” divided by the MUT CPU time. “Productivity” tells you what percentage of the Total CPU and wall clock elapsed times are spent in the mutator (MUT).
-Sflag, as well as giving the same output as the
-sflag, prints information about each GC as it happens:
Alloc Copied Live GC GC TOT TOT Page Flts bytes bytes bytes user elap user elap 528496 47728 141512 0.01 0.02 0.02 0.02 0 0 (Gen: 1) [...] 524944 175944 1726384 0.00 0.00 0.08 0.11 0 0 (Gen: 0)
For each garbage collection, we print:
- How many bytes we allocated this garbage collection.
- How many bytes we copied this garbage collection.
- How many bytes are currently live.
- How long this garbage collection took (CPU time and elapsed wall clock time).
- How long the program has been running (CPU time and elapsed wall clock time).
- How many page faults occurred this garbage collection.
- How many page faults occurred since the end of the last garbage collection.
- Which generation is being garbage collected.
5.7.5. RTS options for concurrency and parallelism¶
The RTS options related to concurrency are described in Using Concurrent Haskell, and those for parallelism in RTS options for SMP parallelism.
5.7.6. RTS options for profiling¶
Most profiling runtime options are only available when you compile your program for profiling (see Compiler options for profiling, and RTS options for heap profiling for the runtime options). However, there is one profiling option that is available for ordinary non-profiled executables:
Generates a basic heap profile, in the file
prog.hp. To produce the heap profile graph, use hp2ps (see hp2ps – Rendering heap profiles to PostScript). The basic heap profile is broken down by data constructor, with other types of closures (functions, thunks, etc.) grouped into broad categories (e.g.
THUNK). To get a more detailed profile, use the full profiling support (Profiling). Can be shortened to
The meaning of the shortened
-his dependent on whether your program was compiled for profiling. (See RTS options for heap profiling for details.)
Default: 25 characters
Sets the maximum length of the cost-centre names listed in the heap profile.
When the program is linked with the
(Options affecting linking), runtime events can be logged in several ways:
- In binary format to a file for later analysis by a variety of tools. One such tool is ThreadScope, which interprets the event log to produce a visual parallel execution profile of the program.
- In binary format to customized event log writer. This enables live analysis of the events while the program is running.
- As text to standard output, for debugging purposes.
Log events in binary format. Without any ⟨flags⟩ specified, this logs a default set of events, suitable for use with tools like ThreadScope.
Per default the events are written to
program.eventlogthough the mechanism for writing event log data can be overridden with a custom EventLogWriter.
For some special use cases you may want more control over which events are included. The ⟨flags⟩ is a sequence of zero or more characters indicating which classes of events to log. Currently these the classes of events that can be enabled/disabled:
s— scheduler events, including Haskell thread creation and start/stop events. Enabled by default.
g— GC events, including GC start/stop. Enabled by default.
n— non-moving garbage collector (see
--nonmoving-gc) events including start and end of the concurrent mark and census information to characterise heap fragmentation. Disabled by default.
p— parallel sparks (sampled). Enabled by default.
f— parallel sparks (fully accurate). Disabled by default.
ticky-ticky profilerevents. Disabled by default.
u— user events. These are events emitted from Haskell code using functions such as
Debug.Trace.traceEvent. Enabled by default.
You can disable specific classes, or enable/disable all classes at once:
a— enable all event classes listed above
-⟨x⟩— disable the given class of events, for any event class listed above
-a— disable all classes
-l-agwould disable all event classes (
-a) except for GC events (
For spark events there are two modes: sampled and fully accurate. There are various events in the life cycle of each spark, usually just creating and running, but there are some more exceptional possibilities. In the sampled mode the number of occurrences of each kind of spark event is sampled at frequent intervals. In the fully accurate mode every spark event is logged individually. The latter has a higher runtime overhead and is not enabled by default.
The format of the log file is described in this users guide in Eventlog encodings It can be parsed in Haskell using the ghc-events library. To dump the contents of a
.eventlogfile as text, use the tool
ghc-events showthat comes with the ghc-events package.
Each event is associated with a timestamp which is the number of nanoseconds since the start of executation of the running program. This is the elapsed time, not the CPU time.
Sets the destination for the eventlog produced with the
Default: disabled Since: 9.2
When enabled, the eventlog will be flushed periodically every ⟨seconds⟩. This can be useful in live-monitoring situations where the eventlog is consumed in real-time by another process.
Log events as text to standard output, instead of to the
.eventlogfile. The ⟨flags⟩ are the same as for
-l, with the additional option
twhich indicates that the each event printed should be preceded by a timestamp value (in the binary
.eventlogfile, all events are automatically associated with a timestamp).
The debugging options
-Dx also generate events which are logged
using the tracing framework. By default those events are dumped as text
to stdout (
-v), but they may instead be stored in
the binary eventlog file by using the
5.7.8. RTS options for hackers, debuggers, and over-interested souls¶
These RTS options might be used (a) to avoid a GHC bug, (b) to see “what’s really happening”, or (c) because you feel like it. Not recommended for everyday use!
Sound the bell at the start of each garbage collection.
Oddly enough, people really do use this option! Our pal in Durham (England), Paul Callaghan, writes: “Some people here use it for a variety of purposes—honestly!—e.g., confirmation that the code/machine is doing something, infinite loop detection, gauging cost of recently added code. Certain people can even tell what stage [the program] is in by the beep pattern. But the major use is for annoying others in the same office…”
An RTS debugging flag; only available if the program was linked with the
-debugoption. Various values of ⟨x⟩ are provided to enable debug messages and additional runtime sanity checks in different subsystems in the RTS, for example
+RTS -Ds -RTSenables debug messages from the scheduler. Use
+RTS -?to find out which debug flags are supported.
Full list of currently supported flags:
DEBUG: zero freed memory on GC¶
DEBUG: stack squeezing¶
DEBUG: program coverage¶
Debug messages will be sent to the binary event log file instead of stdout if the
-l ⟨flags⟩option is added. This might be useful for reducing the overhead of debug tracing.
To figure out what exactly they do, the least bad way is to grep the rts/ directory in the ghc code for macros like
Produce “ticky-ticky” statistics at the end of the program run (only available if the program was linked with
-debug). The ⟨file⟩ business works just like on the
-S [⟨file⟩]RTS option, above.
For more information on ticky-ticky profiling, see Using “ticky-ticky” profiling (for implementors).
(Only available when the program is compiled for profiling.) When an exception is raised in the program, this option causes a stack trace to be dumped to
This can be particularly useful for debugging: if your program is complaining about a
head error and you haven’t got a clue which bit of code is causing it, compiling with
-prof) and running with
+RTS -xc -RTSwill tell you exactly the call stack at the point the error was raised.
The output contains one report for each exception raised in the program (the program might raise and catch several exceptions during its execution), where each report looks something like this:
*** Exception raised (reporting due to +RTS -xc), stack trace: GHC.List.CAF --> evaluated by: Main.polynomial.table_search, called from Main.polynomial.theta_index, called from Main.polynomial, called from Main.zonal_pressure, called from Main.make_pressure.p, called from Main.make_pressure, called from Main.compute_initial_state.p, called from Main.compute_initial_state, called from Main.CAF ...
The stack trace may often begin with something uninformative like
GHC.List.CAF; this is an artifact of GHC’s optimiser, which lifts out exceptions to the top-level where the profiling system assigns them to the cost centre “CAF”. However,
+RTS -xcdoesn’t just print the current stack, it looks deeper and reports the stack at the time the CAF was evaluated, and it may report further stacks until a non-CAF stack is found. In the example above, the next stack (after
--> evaluated by) contains plenty of information about what the program was doing when it evaluated
Implementation details aside, the function names in the stack should hopefully give you enough clues to track down the bug.
See also the function
traceStackin the module
Debug.Tracefor another way to view call stacks.
Turn off update frame squeezing on context switch. (There’s no particularly good reason to turn it off, except to ensure the accuracy of certain data collected regarding thunk entry counts.)
5.7.9. Getting information about the RTS¶
It is possible to ask the RTS to give some information about itself. To do this, use the
$ ./a.out +RTS --info [("GHC RTS", "YES") ,("GHC version", "6.7") ,("RTS way", "rts_p") ,("Host platform", "x86_64-unknown-linux") ,("Host architecture", "x86_64") ,("Host OS", "linux") ,("Host vendor", "unknown") ,("Build platform", "x86_64-unknown-linux") ,("Build architecture", "x86_64") ,("Build OS", "linux") ,("Build vendor", "unknown") ,("Target platform", "x86_64-unknown-linux") ,("Target architecture", "x86_64") ,("Target OS", "linux") ,("Target vendor", "unknown") ,("Word size", "64") ,("Compiler unregisterised", "NO") ,("Tables next to code", "YES") ,("Flag -with-rtsopts", "") ]
The information is formatted such that it can be read as a of type
[(String, String)]. Currently the following fields are present:
- Is this program linked against the GHC RTS? (always “YES”).
- The version of GHC used to compile this program.
- The variant (“way”) of the runtime. The most common values are
rts_thr(threaded runtime, i.e. linked using the
rts_p(profiling runtime, i.e. linked using the
-profoption). Other variants include
dyn(the RTS is linked in dynamically, i.e. a shared library, rather than statically linked into the executable itself). These can be combined, e.g. you might have
- These are the platform the program is compiled to run on.
- These are the platform where the program was built on. (That is, the target platform of GHC itself.) Ordinarily this is identical to the target platform. (It could potentially be different if cross-compiling.)
- These are the platform where GHC itself was compiled. Again, this would normally be identical to the build and target platforms.
"64", reflecting the word size of the target platform.
- Was this program compiled with an “unregistered” version of GHC? (I.e., a version of GHC that has no platform-specific optimisations compiled in, usually because this is a currently unsupported platform.) This value will usually be no, unless you’re using an experimental build of GHC.
Tables next to code
- Putting info tables directly next to entry code is a useful performance optimisation that is not available on all platforms. This field tells you whether the program has been compiled with this optimisation. (Usually yes, except on unusual platforms.)
- The value of the GHC flag
-with-rtsopts=⟨opts⟩at compile/link time.