| 1 |
= Memory Management in GNU APL |
| 2 |
:Author: Jürgen Sauermann, GNU APL |
| 3 |
:page-width: 80em |
| 4 |
|
| 5 |
== Abstract |
| 6 |
|
| 7 |
This document describes how memory is being managed in GNU APL and what |
| 8 |
the consequences are for a GNU APL programmer. |
| 9 |
|
| 10 |
== Fundamentals |
| 11 |
|
| 12 |
* The GNU APL interpreter is an executable that runs in a single process. |
| 13 |
Inside the interpreter run a handful of threads; the memory overhead of |
| 14 |
these threads can be neglected. |
| 15 |
|
| 16 |
* Depending on configuration details and user preferences, the interpreter may |
| 17 |
fork additional processes (APserver, GTK_server) whose sizes are so small |
| 18 |
they can also be neglected. |
| 19 |
|
| 20 |
The sizes of the binaries mentioned (at the time of this writing) are: |
| 21 |
|
| 22 |
.Binary Sizes |
| 23 |
[options="header"] |
| 24 |
[width="60%"] |
| 25 |
|=============== |
| 26 |
| Binary | Size (bytes) |
| 27 |
|GNU APL Interpreter | 44,620,848 |
| 28 |
|APserver | 1,540,944 |
| 29 |
|Gtk_server | 530,376 |
| 30 |
|=============== |
| 31 |
|
| 32 |
* After the interpreter has started it requests additional memory from the |
| 33 |
operating system on behalf of APL commands and APL code executed. That |
| 34 |
memory is allocated by malloc(), either directly or indirectly (via Cpp |
| 35 |
new()). When the memory is no longer needed, it is returned to malloc() |
| 36 |
memory management (for later re-use). The memory kept in the malloc() |
| 37 |
memory management is only returned to the operating system when the |
| 38 |
GNU APL interpreter exits. |
| 39 |
|
| 40 |
* GNU APL tries to allocate memory statically as much as possible. The |
| 41 |
static allocation has already taken place when the interpreter outputs its |
| 42 |
first APL prompt. After that dynamic memory allocation via malloc()/new() |
| 43 |
is performed for objects whose lifetime is potentially infinite (i.e. whose |
| 44 |
destruction is triggered by APL code. These objects are: |
| 45 |
** APL variables |
| 46 |
** APL values |
| 47 |
** defined functions |
| 48 |
** call stack (aka. *)SI* stack) entries of defined functions |
| 49 |
|
| 50 |
In the following the sizes of these objects are given. The sizes differ |
| 51 |
slightly between 32-bit and 64-bit systems because 32-bit systems use 4-byte |
| 52 |
pointers and 64-bit systems use 8-byte pointers. In addition to the indicated |
| 53 |
sizes, there is a (very) small overhead in the malloc() library for every |
| 54 |
object. Each of the objects is significantly larger than its malloc overhead |
| 55 |
and therefore the overhead can normally be ignored. A typical APL workspace is |
| 56 |
assumed to have fewer than 1000 APL symbols (i.e. variables or defined |
| 57 |
functions) and an )SI stack depth of less than 100. A typical computer on which |
| 58 |
GNU APL runs is assumed to have 4 GB of memory available to the interpreter and |
| 59 |
it should definitely have at least 1 GB. |
| 60 |
|
| 61 |
.Sizes of Dynamic Objects |
| 62 |
[options="header"] |
| 63 |
[width="60%"] |
| 64 |
|=============== |
| 65 |
| Object | 32-bit system | 64-bit system |
| 66 |
| ValueStackItem | 16 | 24 |
| 67 |
| Token | 20 | 24 |
| 68 |
| Ravel Cell | 20 | 24 |
| 69 |
| UCS_string | 20 + X | 24 + X |
| 70 |
| APL Symbol | 56 + X | 80 + X |
| 71 |
| Defined Function | 292 + X | 400 + X |
| 72 |
| Call Stack Entry | 2952 | 3952 |
| 73 |
| APL Value | 372 + X | 456 + X |
| 74 |
|=============== |
| 75 |
|
| 76 |
In the table above, the term + X means that the object is a structure that |
| 77 |
contains other objects. For example, the size of a defined function depends |
| 78 |
on the number of function lines and their lengths. Likewise, the size of an |
| 79 |
APL value depends primarily on the number of its ravel elements. The term X |
| 80 |
for the different objects is explained below. |
| 81 |
|
| 82 |
=== X for a Unicode string (aka. UCS string) |
| 83 |
|
| 84 |
X is 4 bytes per Unicode character. For performance reasons at least 16 |
| 85 |
characters are allocated so that X ≥ 64. |
| 86 |
|
| 87 |
=== X for an APL Symbol |
| 88 |
|
| 89 |
An APL symbol has two variable parts: a UCS string containing the name of the |
| 90 |
symbol (e.g. a variable name or the name of a defined function) and a value |
| 91 |
stack. The value stack of a function is initially empty, but is increased |
| 92 |
when a defined function is called and the symbol is a local variable or a |
| 93 |
label in the defined function and decreased if that function returns. The |
| 94 |
amount of memory by which the stack is increased is the size of a |
| 95 |
*ValueStackItem* in the table above. |
| 96 |
|
| 97 |
=== X for a Defined Function |
| 98 |
|
| 99 |
When a defined function is created, its textual representation (i.e. the |
| 100 |
argument for ⎕FX or the result produced by the ∇-editor) is stored in the |
| 101 |
function object for the purpose of displaying error information in case the |
| 102 |
execution of the function fails. |
| 103 |
|
| 104 |
The textual representation is then converted into an internal representation |
| 105 |
of fixed-sized Tokens with one token for each lexical unit. A lexical unit is: |
| 106 |
|
| 107 |
* one user-defined name, or |
| 108 |
* one distinguished name, or |
| 109 |
* one primitive function, or |
| 110 |
* a special symbol (◊ [ ] etc.) |
| 111 |
|
| 112 |
Finally, a jump table to the different function lines is added to the |
| 113 |
function. The X for the function is therefore Utx + N×To + L×Jt, where |
| 114 |
|
| 115 |
* Utx is the size of the UCS_string containing the function text, |
| 116 |
* N is the number of Tokens, |
| 117 |
* To is the size of a Token as above, |
| 118 |
* L is the number of function lines |
| 119 |
* Jt is the size of a jump table entry (4 bytes) |
| 120 |
|
| 121 |
=== X for an APL Value |
| 122 |
|
| 123 |
GNU APL distinguishes between short values and long values. A short value is a |
| 124 |
value with up to 12 ravel elements while a long value is a value with more |
| 125 |
than 12 ravel elements. The split point between long and short values is 12 by |
| 126 |
default, but can be changed via ./configure. |
| 127 |
|
| 128 |
The rationale for distinguishing short and long values is performance. A short |
| 129 |
value can be allocated in a single new() call while a long value needs two |
| 130 |
new() calls: one for the value header (which contains the shape of the value), |
| 131 |
and one for the ravel cells of the value. Looking at the details of malloc() |
| 132 |
one can see that the allocation or deallocation of smaller memory areas is more |
| 133 |
efficient than the allocation or deallocation of larger areas, Finally, short |
| 134 |
APL values occur more frequently in APL programs than long values, and they |
| 135 |
are also more frequently overridden as a whole (which allocates a new value |
| 136 |
and discards the old) as opposed to being updated via indexed assignment |
| 137 |
(which changes an existing value but does not create a newly allocated |
| 138 |
value (unless nesting is involved)). |
| 139 |
|
| 140 |
* If a value is short then X = 0 |
| 141 |
* otherwise X = N × Ce, where |
| 142 |
** N is the number of ravel elements (i.e. ⍴,VALUE), and |
| 143 |
** Ce is the size of a ravel Cell as above. |
| 144 |
|
| 145 |
*Note:* In the classical APL world of ISO 8485, all ravel elements of an APL |
| 146 |
value had the same type such as bit, 8-bit char, 32-bit integer, or float and |
| 147 |
the ravel was densely packed. Accessing a particular element V[n] of a value V |
| 148 |
was a fast constant time operation, which was internally performed very |
| 149 |
frequently. Many APL primitive functions (with the noteworthy exception of |
| 150 |
scalar functions) need to perform multiple computations of V[n] for one |
| 151 |
of their arguments V. |
| 152 |
|
| 153 |
In ISO 13751 the concept of mixed values was introduced. In a mixed value, |
| 154 |
different ravel elements can have different types, so that the type becomes a |
| 155 |
property of each ravel element of a value instead of a property of the value |
| 156 |
itself. The implementer of ISO 13751 then had different alternatives (and |
| 157 |
possibly combinations of them) to deal with mixed values: |
| 158 |
|
| 159 |
* Leave the existing bit-, byte-, integer- and float-arrays as they |
| 160 |
are and add a new kind of mixed-arrays. This was a reasonable approach for |
| 161 |
those who had a code base from ISO 8485 already and who wanted to extend |
| 162 |
that code base to ISO 13751. |
| 163 |
|
| 164 |
* Make the ravel elements pointers to a different location that contains the |
| 165 |
real ravel elements. Since all pointers have the same size, the constant-time |
| 166 |
access to arbitrary ravel elements would remain. The performance cost |
| 167 |
would be somewhat higher due to: |
| 168 |
** the additional indirection via the pointers, and |
| 169 |
** possibly one more memory allocation per ravel element. |
| 170 |
While the cost for the additional indirection would probably be more than |
| 171 |
compensated by the reduced access time, the cost for additional memory |
| 172 |
allocations could have become a major headache. |
| 173 |
|
| 174 |
* Make the ravel elements virtual objects that all have the same size. That |
| 175 |
means instead of primitive functions working with pointers to ravel |
| 176 |
elements, the ravel elements would point to the functions that are |
| 177 |
appropriate for their type. On one hand, this approach wastes quite some |
| 178 |
memory because the largest possible type (i.e. complex numbers) determines |
| 179 |
the size of all ravel cells including those that would fit in far fewer |
| 180 |
bytes. On the other hand, there are quite a few advantages: |
| 181 |
** The same size of all elements causes V[n] to take constant time, |
| 182 |
** The indexed assignment of e.g. a single element (of a different type) to a |
| 183 |
non-mixed array would also take (constant time (in option 1 above the entire |
| 184 |
array would have needed to be converted from non-mixed to mixed)). |
| 185 |
** The number of different array types would be reduced to 1 instead of being |
| 186 |
increased from 5 to 6 (bit, byte, integer, float, complex, and mixed). This |
| 187 |
in turn makes a huge number of type checks in the dyadic APL primitives |
| 188 |
obsolete. Since a virtual cell knows its own type, it only needs to check |
| 189 |
the type of the other cell(s) in a dyadic operation. |
| 190 |
** The "wasted" space could be used for other features at no extra cost such as: |
| 191 |
*** 64-bit integers instead of 32-bit Integers, |
| 192 |
*** rational number arithmetic for floating point numbers |
| 193 |
|
| 194 |
Since GNU APL was to be designed from scratch without having a prior code base, |
| 195 |
the first option was ruled out almost immediately. The second option was |
| 196 |
briefly considered, but then ruled out as well because code size for the second |
| 197 |
option was expected to be significantly larger than for the third option. |
| 198 |
|
| 199 |
== Summary |
| 200 |
|
| 201 |
The discussion of sizes so far can be summarized as: |
| 202 |
|
| 203 |
*Corollary:* For a typical APL program on a typical computer, the only items |
| 204 |
of concern in the context of memory management are APL values. |
| 205 |
|
| 206 |
*Note:* The default values of system variable ⎕SYL (aka. system limits) |
| 207 |
guarantee to some extent that an APL program behaves like a typical APL |
| 208 |
program, |
| 209 |
|
| 210 |
== How the Available Memory is Estimated |
| 211 |
|
| 212 |
=== History |
| 213 |
|
| 214 |
In the good old days of CP/M and friends, say 1975, memory management was |
| 215 |
rather simple. The process running an APL interpreter would know how much |
| 216 |
memory was reserved (and therefore available for it) and it could use that |
| 217 |
memory without restrictions. The amount of memory (say *initial-⎕WA*) that |
| 218 |
the process could use was simply: |
| 219 |
|
| 220 |
* *initial-⎕WA = top-of-memory - top-of-used-memory* |
| 221 |
|
| 222 |
This value was determined at the start-up of the interpreter (where the malloc() |
| 223 |
pool was essentially empty) and the interpreter would keep track of the memory |
| 224 |
that it allocated after that, say *memory-used*. So printing ⎕WA at any time |
| 225 |
would essentially show *initial-⎕WA - memory-used*. And that was it. |
| 226 |
|
| 227 |
=== Present Memory Management |
| 228 |
|
| 229 |
Since then a number of things have changed. The changes are in general |
| 230 |
advantageous, but also cause trouble in very specific situations. The main |
| 231 |
differences, as seen by an application like GNU APL are: |
| 232 |
|
| 233 |
* Change from physical to virtual memory. These days the memory allocated by |
| 234 |
malloc() is virtual. Under *normal* circumstances the application need not |
| 235 |
care about this difference, but in some special cases it is negatively |
| 236 |
affected (and can then do very little about it). |
| 237 |
|
| 238 |
* A direct consequence of virtual memory is that the size of it is no longer |
| 239 |
determined by the physical memory available in a computer, but by other |
| 240 |
properties like the number of bits in the virtual memory addresses. This |
| 241 |
makes it impossible to determine initial-⎕WA (more precisely: ⎕WA is no |
| 242 |
longer a constant but becomes dependent on what happens not only in the APL |
| 243 |
interpreter itself but also in other processes and/or in the kernel. |
| 244 |
|
| 245 |
* Over-commitment: Modern operating systems return far more virtual memory to |
| 246 |
applications than they really have in terms of physical memory. The idea is |
| 247 |
that not all applications reach their maximum need for memory at the same |
| 248 |
time, so that the same piece of physical memory can be used by different |
| 249 |
applications (or by the kernel for that matter) at different times. |
| 250 |
|
| 251 |
*Corollary 1:* In the old days, the memory that an application has obtained |
| 252 |
from the kernel was a *guarantee*. These days it is merely a *promise* which |
| 253 |
is *normally* kept but may fail with a small though non-zero probability. |
| 254 |
|
| 255 |
*Corollary 2:* As far as GNU APL is concerned, GNU APL does not (and actually |
| 256 |
cannot) guarantee proper operation if the kernel reaches the limit of its |
| 257 |
physical memory. Some error handling mechanisms, in particular WS FULL errors, |
| 258 |
may fail in a non-graceful fashion - including immediate termination of the |
| 259 |
GNU APL process without any warnings. |
| 260 |
|
| 261 |
Although a crash of GNU APL cannot be prevented in general, one can decrease |
| 262 |
the probability of such crashes by taking countermeasures beforehand. Most GNU |
| 263 |
APL users will not need this (see *normal* above), but those facing problems |
| 264 |
with improper WS FULL handling should read on. |
| 265 |
|
| 266 |
== Improving the GNU APL WS FULL handling |
| 267 |
|
| 268 |
=== The OOM handler |
| 269 |
|
| 270 |
One enemy of GNU APL is the kernel's OOM (out-of-memory) handler. The OOM |
| 271 |
handler is invoked when the kernel needs (physical) memory and has none. It |
| 272 |
then kills one or more running processes in order to claim their memory back. |
| 273 |
The processes that are killed (ungracefully) may or may not be the process(es) |
| 274 |
that have caused the shortage of memory. That is: |
| 275 |
|
| 276 |
* GNU APL may be killed by some other (unrelated) process or interrupt that |
| 277 |
has requested memory, or |
| 278 |
* some other (unrelated) process may be killed by a memory request from GNU |
| 279 |
APL, or |
| 280 |
* GNU APL is killed as a result of its own memory request. |
| 281 |
|
| 282 |
It should be clear that the first two cases are very difficult to reproduce. |
| 283 |
The third case is much easier by creating a very large APL value. It is also |
| 284 |
the case that has been observed in reality. |
| 285 |
|
| 286 |
In GNU/Linux, one can protect individual processes from being killed by the |
| 287 |
OOM handler with the following command (as root; <PID> is the process ID of |
| 288 |
the process to be protected: |
| 289 |
|
| 290 |
---- |
| 291 |
echo -17 > /proc/<PID>/oom_adj |
| 292 |
---- |
| 293 |
|
| 294 |
The OOM handler can also be disabled entirely with |
| 295 |
|
| 296 |
---- |
| 297 |
sysctl vm.overcommit_memory=2 |
| 298 |
---- |
| 299 |
|
| 300 |
or: |
| 301 |
|
| 302 |
---- |
| 303 |
echo "vm.overcommit_memory=2" >> /etc/sysctl.conf |
| 304 |
---- |
| 305 |
|
| 306 |
*Warning:* These settings can seriously impact the stability of the operating |
| 307 |
system and should not be used on machines whose availability is a concern. |
| 308 |
There is no point in sacrificing the stability of a system for the stability of |
| 309 |
a process running on that system. |
| 310 |
|
| 311 |
=== Using System Limits |
| 312 |
|
| 313 |
Instead of disabling the OOM handler, one should use methods that prevent GNU |
| 314 |
APL from requesting too much memory, so that the OOM handler is not invoked |
| 315 |
in the first place (at least not from GNU APL). There are several ⎕SYL (aka. |
| 316 |
system limit) entries that can be set by an APL program to prevent GNU APL |
| 317 |
from requesting too much memory. Some of them have to be combined with others |
| 318 |
to achieve full protection. |
| 319 |
|
| 320 |
* ⎕SYL[1;] limits the depth of the )SI stack and therefore the amount of |
| 321 |
memory spent for function calls. Referring to Table 2 above, *⎕SYL[1;2]←1000* |
| 322 |
will throw a LIMIT ERROR if the )SI stack grows above 3 Megabytes (32-bit |
| 323 |
systems) or 4 Megabytes (64-bit systems). |
| 324 |
* ⎕SYL[2;] limits the number of APL values. *⎕SYL[2;2]←1000* will throw a |
| 325 |
LIMIT ERROR when more than 1000 APL values (variables as well as |
| 326 |
intermediate results) are created. Referring to Table 2 above, this |
| 327 |
corresponds to 372 kByte on 32-bit systems, or 456 kByte on 64-bit systems. |
| 328 |
Note that the space for localized variables of defined functions belongs to |
| 329 |
⎕SYL[2;] and not to ⎕SYL[1;]. |
| 330 |
* ⎕SYL[3;] limits the number of ravel elements in APL values. |
| 331 |
*⎕SYL[3;2]←1000000* will throw a LIMIT ERROR before more than 20 MByte on |
| 332 |
32-bit systems, or more than 24 MByte on 64-bit systems is requested. |
| 333 |
* Finally, ⎕SYL[31 32;] define a safety margin that is is explained below. |
| 334 |
|
| 335 |
In short, ⎕SYL[1;] protects primarily against too deep recursion, which is most |
| 336 |
likely caused by a programming fault. ⎕SYL[2;] protects against too many APL |
| 337 |
values, and ⎕SYL[3;] protects against APL values becoming too large. |
| 338 |
|
| 339 |
Like WS FULL and other APL errors, hitting a system limit brings the |
| 340 |
interpreter back to immediate execution (interactive) mode. If one wants to |
| 341 |
handle the situation programmatically, one can catch the limit error with ⎕EA, |
| 342 |
⎕EB, or ⎕EC. |
| 343 |
|
| 344 |
|
| 345 |
=== Analyzing /proc/meminfo |
| 346 |
|
| 347 |
On a GNU/Linux machine, the file */proc/meminfo* provides quite useful |
| 348 |
information about the usage of memory at the point in time when /proc/meminfo |
| 349 |
is read. Before setting system limits as described above, one should consult |
| 350 |
/proc/meminfo to see how the memory is distributed and then set the limits |
| 351 |
accordingly. |
| 352 |
|
| 353 |
Ideally */proc/meminfo* shows something like: |
| 354 |
|
| 355 |
---- |
| 356 |
MemTotal: 185736 kB |
| 357 |
MemFree: 12660 kB |
| 358 |
MemAvailable: 126264 kB |
| 359 |
Buffers: 74704 kB |
| 360 |
Cached: 54732 kB |
| 361 |
---- |
| 362 |
|
| 363 |
In that case *MemAvailable:* is the value on which the system limits settings |
| 364 |
should be based. On older GNU/Linux machines *MemAvailable:* may not be |
| 365 |
displayed. In that case *MemFree: + Cached:* can be used as a workaround. |
| 366 |
|
| 367 |
The actual implementation of *MemAvailable:* by the kernel is quite complicated |
| 368 |
(more than simply adding *MemFree:* and *Cached:*). For that reason, even if |
| 369 |
*MemAvailable:* were available on all GNU/Linux machines it would still be |
| 370 |
impossible to call it before each creation of an APL value. |
| 371 |
|
| 372 |
Instead, GNU APL uses the following simple algorithm: |
| 373 |
|
| 374 |
* After the interpreter has started, compute *total_memory* as follows: |
| 375 |
** If the process that runs the interpreter has a limit set on the amount of |
| 376 |
its virtual memory, then the interpreter assumes that amount of memory |
| 377 |
will be available and sets *total_memory* to that limit. In this case |
| 378 |
*/proc/meminfo* is not used. |
| 379 |
** Otherwise (i.e. the virtual memory for the process is unlimited) the |
| 380 |
interpreter consults */proc/meminfo* and sets *total_memory* to either |
| 381 |
*MemAvailable:* (if present) or else to *MemFree: + Cached:*. |
| 382 |
* As new APL values are created and destroyed in the course of running APL |
| 383 |
programs, the amount of memory allocated for them is tracked in, say, |
| 384 |
*used_memory*. |
| 385 |
* Before requesting new memory of, say, size *new*, GNU APL checks that |
| 386 |
|
| 387 |
---- |
| 388 |
total_memory_1 ≥ used_memory + new + ⎕WA-margin |
| 389 |
---- |
| 390 |
* In this check *total_memory_1* is *⎕SYL[32;2]* percent of *total_memory* |
| 391 |
above, and reduced by *⎕SYL[31;2]*. The reason for the scaling by |
| 392 |
*⎕SYL[32;2]* is that malloc() usually requests slightly more memory from |
| 393 |
the kernel than the application had requested from malloc(). |
| 394 |
|
| 395 |
** If that check fails then WS FULL is raised, |
| 396 |
** otherwise the request is forwarded to malloc() and from there possibly |
| 397 |
to the kernel. |
| 398 |
|
| 399 |
The ⎕WA-margin currently has a default value of 0, but can be changed via |
| 400 |
|
| 401 |
=== Hints for GNU APL Users |
| 402 |
|
| 403 |
==== Setting a memory limit |
| 404 |
|
| 405 |
If stability of a system is a concern, then setting memory limits for |
| 406 |
processes is normally a good idea. The command for doing that differs |
| 407 |
between shells; in *bash* you can start GNU APL with a virtual memory limit |
| 408 |
of like this: |
| 409 |
|
| 410 |
---- |
| 411 |
ulimit -v 1000000 |
| 412 |
apl |
| 413 |
---- |
| 414 |
The unit is kB, so that the command above will set the limit to 1 GB. The limit |
| 415 |
remains in force for all subsequent commands and can be removed like this: |
| 416 |
|
| 417 |
---- |
| 418 |
ulimit -v unlimited |
| 419 |
apl |
| 420 |
---- |
| 421 |
|
| 422 |
*Note:* setting memory limits via *ulimit* as such does not directly solve the |
| 423 |
problem that a process may have when it reaches the limit. However, it helps a |
| 424 |
process to predict when the limit will be reached. |
| 425 |
|
| 426 |
=== 32-bit GNU/Linux |
| 427 |
|
| 428 |
On 32-bit machines, there is a per process memory limit of about 3 GByte. If |
| 429 |
the memory of a machine is, say, 8 GB or more, then one can have several |
| 430 |
processes with each process allocating up to 3 GByte as long as the memory |
| 431 |
shown in */proc/meminfo* is not exceeded. |
| 432 |
|
| 433 |
If GNU APL sees a *total_memory* of more than 3 GB in the analysis of |
| 434 |
*/proc/meminfo*, then it reduces *total_memory* to about 3 GB. if it is |
| 435 |
running on a 32-bit kernel. 64-bit machines have a similar limit, but the |
| 436 |
limit is so high that it is not relevant in practice. |
| 437 |
|
| 438 |
The safety margin *(total_memory_1* - *total_memory)* that results from |
| 439 |
*⎕SYL[31 32;2]* remains in effect so that the total memory that an APL program |
| 440 |
can obtain is somewhere below 3 GB. |
| 441 |
|
| 442 |
==== Troubleshooting WS FULL problems |
| 443 |
|
| 444 |
For debugging purposes the safety margin can be disabled in APL like this: |
| 445 |
|
| 446 |
---- |
| 447 |
⎕SYL[31;2] ← 1000000 ⍝ smallest margin (1 MB) |
| 448 |
⎕SYL[32;2] ← 200 ⍝ far above available |
| 449 |
---- |
| 450 |
|
| 451 |
WS FULL can occur very many places within the interpreter. To find the exact |
| 452 |
position where a WS FULL was generated, some related logging facilities can be |
| 453 |
turned on: |
| 454 |
|
| 455 |
---- |
| 456 |
]log 25 |
| 457 |
]log 26 |
| 458 |
]log 45 |
| 459 |
---- |
| 460 |
On a 32-bit machine a WS FULL can be reliably produced like this (the example |
| 461 |
was kindly provided by Christian Robert): |
| 462 |
|
| 463 |
---- |
| 464 |
N←5000000 ⍝ 32-bit; use a larger value for 64-bit |
| 465 |
A←⊂'0123456789ABCDEF' |
| 466 |
↑↑ N ⍴ A ⍝ trigger WS FULL |
| 467 |
---- |
| 468 |
This example is of particular interest because, depending upon variable *N* |
| 469 |
it creates two different challenges for the kernel. One is a big allocation |
| 470 |
when the process is already near starvation and one is a small allocation |
| 471 |
that fails. |
| 472 |
|
| 473 |
In both cases the WS FULL is caused by *N ⍴ A*. Because A is a scalar (due to |
| 474 |
⊂), all ravel elements of the result are nested values. Therefore the |
| 475 |
computation of *N ⍴ A* creates exactly 1+N values: a large value for the result |
| 476 |
itself and N small nested values - the ravel of the result. |
| 477 |
|
| 478 |
Referring to Table 2, the variable *A* (and therefore each nested value of the |
| 479 |
result needs *372 + 16×20 = 692* bytes. The top-level of the result requires |
| 480 |
*372 + N×20* bytes. Let *F* be the amount of (free) memory that could be |
| 481 |
obtained from malloc(). Then: |
| 482 |
|
| 483 |
1. if *F < 372 + N×20* then the allocation of the top-level value fails, |
| 484 |
|
| 485 |
2. else if *F < 372 + N×20 + N×692* then the allocation of the top-level value |
| 486 |
succeeds, but the allocation of one of the nested sub-values fails, |
| 487 |
|
| 488 |
3. else all allocations succeed and *N ⍴ A* succeeds as expected. |
| 489 |
|
| 490 |
The less than obvious difference between the first and the second case is the |
| 491 |
amount of free memory at the point in time where the allocation failed: |
| 492 |
|
| 493 |
1. if the first (large) allocation fails, then the free memory is still F, |
| 494 |
|
| 495 |
2. if one of the subsequent (small) allocations fails then (since the previous |
| 496 |
allocations have succeeded) the free memory is now less than the requested |
| 497 |
size, regardless how large F was initially. In our example F would be less |
| 498 |
than 692 bytes (in theory) or less than about 1024 (in practice because |
| 499 |
malloc allocates the next power of 2). |
| 500 |
|
| 501 |
Exception handling with only 1024 bytes left can fail as can be demonstrated |
| 502 |
on a 32-bit machine quite easily. The following examples show both cases (with |
| 503 |
safety margin disabled in order to allow near-memory-starvation. |
| 504 |
|
| 505 |
==== Example 1: WS FULL with sufficient free memory left |
| 506 |
|
| 507 |
---- |
| 508 |
|
| 509 |
⍝ enable relevant logging facilities |
| 510 |
]log 25 |
| 511 |
Log facility 'more verbose errors ' is now ON |
| 512 |
]log 26 |
| 513 |
Log facility 'details of error throwing ' is now ON |
| 514 |
]log 45 |
| 515 |
Log facility 'details of Value allocation ' is now ON |
| 516 |
|
| 517 |
⍝ disable safety margin |
| 518 |
⎕SYL[31 32;2]←1000000 200 |
| 519 |
⎕SYL[31 32;] |
| 520 |
⎕WA safety margin (bytes) 1000000 |
| 521 |
⎕WA memory scale (%) 200 |
| 522 |
|
| 523 |
A←⊂'0123456789ABCDEF' ⍝ 692 bytes |
| 524 |
N←200000000 ⍝ for 200 Mio ravel Cells |
| 525 |
↑↑ N⍴A ⍝ 4 GB top-level + 138.4 GB nested sub-values |
| 526 |
|
| 527 |
throwing WS FULL at PrimitiveFunction.cc:230 |
| 528 |
|
| 529 |
---------------------------------------- |
| 530 |
-- Stack trace at Error.cc:184 |
| 531 |
---------------------------------------- |
| 532 |
0xB7160AF3 __libc_start_main |
| 533 |
0x8092998 main |
| 534 |
0x8247A9D Workspace::immediate_execution(bool) |
| 535 |
0x80EF330 Command::process_line() |
| 536 |
0x80EFCAB Command::do_APL_expression(UCS_string&) |
| 537 |
0x80EF459 Command::finish_context() |
| 538 |
0x80FB64D Executable::execute_body() const |
| 539 |
0x81DBE31 StateIndicator::run() |
| 540 |
0x8149E2C Prefix::reduce_statements() |
| 541 |
0x8146913 Prefix::reduce_A_F_B_() |
| 542 |
0x815579D Bif_F12_RHO::eval_AB(Value_P, Value_P) |
| 543 |
0x8155463 Bif_F12_RHO::do_reshape(Shape const&, Value const&) |
| 544 |
0x80B5EE6 Value_P::Value_P(Shape const&, char const*) |
| 545 |
0x8245B9A Value::init_ravel() |
| 546 |
0x80FA35B throw_apl_error(ErrorCode, char const*) |
| 547 |
======================================== |
| 548 |
WS FULL+ |
| 549 |
↑↑N⍴A |
| 550 |
^ ^ |
| 551 |
Ravel allocation failed |
| 552 |
|
| 553 |
)MORE |
| 554 |
new Value(PrimitiveFunction.cc:230) failed (APL error in ravel allocation) |
| 555 |
throwing WS FULL at Value_P.icc:227 |
| 556 |
---- |
| 557 |
|
| 558 |
==== Example 2: WS FULL with very little free memory left |
| 559 |
|
| 560 |
---- |
| 561 |
⍝ enable relevant logging facilities |
| 562 |
]log 25 |
| 563 |
Log facility 'more verbose errors ' is now ON |
| 564 |
]log 26 |
| 565 |
Log facility 'details of error throwing ' is now ON |
| 566 |
]log 45 |
| 567 |
Log facility 'details of Value allocation ' is now ON |
| 568 |
|
| 569 |
⍝ disable safety margin |
| 570 |
⎕SYL[31 32;2]←1000000 200 |
| 571 |
⎕SYL[31 32;] |
| 572 |
⎕WA safety margin (bytes) 1000000 |
| 573 |
⎕WA memory scale (%) 200 |
| 574 |
|
| 575 |
A←⊂'0123456789ABCDEF' ⍝ 692 bytes |
| 576 |
N←20000000 ⍝ for 20 Mio ravel Cells |
| 577 |
↑↑ N⍴A ⍝ 0,4 GB top-level + 13.84 GB nested sub-values |
| 578 |
Value_P::Value_P(const Shape & shape, const char * loc) failed at |
| 579 |
Value_P.icc:235 (caller: PrimitiveFunction.cc:240) |
| 580 |
what: std::bad_alloc |
| 581 |
initial sbrk(): 0xa28a000 |
| 582 |
current sbrk(): 0x9eaec000 |
| 583 |
alloc_size: 0x140 (320) |
| 584 |
used memory: 0xbbb495c0 (3149174208) |
| 585 |
|
| 586 |
throwing WS FULL at Value.cc:233 |
| 587 |
|
| 588 |
---------------------------------------- |
| 589 |
-- Stack trace at Error.cc:184 |
| 590 |
---------------------------------------- |
| 591 |
backtrace_symbols() failed. Using backtrace_symbols_fd() instead... |
| 592 |
|
| 593 |
./apl[0x809a380] |
| 594 |
/lib/i386-linux-gnu/libc.so.6(__libc_start_main+0xf3)[0xb70c0af3] |
| 595 |
./apl(main+0x48)[0x8092998] |
| 596 |
./apl(_ZN9Workspace19immediate_executionEb+0x1d)[0x8247a9d] |
| 597 |
./apl(_ZN7Command12process_lineEv+0x460)[0x80ef330] |
| 598 |
./apl(_ZN7Command17do_APL_expressionER10UCS_string+0x6b)[0x80efcab] |
| 599 |
./apl(_ZN7Command14finish_contextEv+0x29)[0x80ef459] |
| 600 |
./apl(_ZNK10Executable12execute_bodyEv+0x1d)[0x80fb64d] |
| 601 |
./apl(_ZN14StateIndicator3runEv+0x21)[0x81dbe31] |
| 602 |
./apl(_ZN6Prefix17reduce_statementsEv+0x22c)[0x8149e2c] |
| 603 |
./apl(_ZN6Prefix13reduce_A_F_B_Ev+0xd3)[0x8146913] |
| 604 |
./apl(_ZN11Bif_F12_RHO7eval_ABE7Value_PS0_+0xad)[0x815579d] |
| 605 |
./apl(_ZN11Bif_F12_RHO10do_reshapeERK5ShapeRK5Value+0x258)[0x8155678] |
| 606 |
./apl(_ZNK11PointerCell10init_otherEPvR5ValuePKc+0x49)[0x814c6e9] |
| 607 |
./apl(_ZNK5Value5cloneEPKc+0x30)[0x82438a0] |
| 608 |
./apl(_ZN7Value_PC1ERK5ShapePKc+0x13e)[0x80b5fae] |
| 609 |
./apl(_ZN5Value15catch_exceptionERKSt9exceptionPKcS4_S4_+0x31e)[0x823c93e] |
| 610 |
./apl(_Z15throw_apl_error9ErrorCodePKc+0x6b)[0x80fa35b] |
| 611 |
======================================== |
| 612 |
WS FULL+ |
| 613 |
↑↑N⍴A |
| 614 |
^ ^ |
| 615 |
|
| 616 |
)MORE |
| 617 |
new Value(const Shape & shape, const char * loc) failed (std::bad_alloc) |
| 618 |
throwing WS FULL at Value_P.icc:227 |
| 619 |
---- |
| 620 |
|
| 621 |
==== Differences between Example 1 and Example 2 |
| 622 |
|
| 623 |
* Execution time: the second example uses noticeably more time before the WS |
| 624 |
FULL occurs. In this time the interpreter creates a large number of |
| 625 |
sub-values until the memory is exhausted. |
| 626 |
* The stack dumps differ. The pretty-printing of Cpp function names requires |
| 627 |
extra memory, in example 2 that extra memory is not available so that |
| 628 |
backtrace_symbols() fails (and the interpreter then uses a function that |
| 629 |
needs less memory): |
| 630 |
|
| 631 |
---- |
| 632 |
---------------------------------------- |
| 633 |
-- Stack trace at Error.cc:184 |
| 634 |
---------------------------------------- |
| 635 |
backtrace_symbols() failed. Using backtrace_symbols_fd() instead... |
| 636 |
|
| 637 |
---- |
| 638 |
|
| 639 |
* The WS FULL is thrown from different locations. PrimitiveFunction.cc:230 is |
| 640 |
in the implementation of dyadic ⍴, while Value.cc:233 is the exception |
| 641 |
handler for *std::bad_alloc* (see also *what:* in the debug printout), |
| 642 |
|
| 643 |
---- |
| 644 |
throwing WS FULL at PrimitiveFunction.cc:230 |
| 645 |
---- |
| 646 |
---- |
| 647 |
throwing WS FULL at Value.cc:233 |
| 648 |
---- |
| 649 |
|
| 650 |
* The )MORE information in the two examples differ. The term *Ravel |
| 651 |
allocation* in the printout of example 1 refers to the allocation of the |
| 652 |
(large) top-level ravel. |
| 653 |
|
| 654 |
=== Final Remarks |
| 655 |
|
| 656 |
The exact behavior of processes with too little memory is generally difficult |
| 657 |
to predict. While the A⍴N example given above is understood to some extent, |
| 658 |
there are other cases such as out-of-memory when the process stack is |
| 659 |
exhausted that were not yet reported. |
| 660 |
|
| 661 |
We are trying our best to fix such cases, but when reporting them as errors |
| 662 |
on bug-apl please consider the following. |
| 663 |
|
| 664 |
* It helps a lot if you can provide a reliable way to reproduce the problem. |
| 665 |
Most of the time for fixing an error is spent on reproducing it. |
| 666 |
* Check that the problem cannot be fixed by setting memory limits. As explained |
| 667 |
above there are some boundary conditions from the operating system that |
| 668 |
cannot be fixed (or even properly handled) by a process. Remember the |
| 669 |
existing tools for that: |
| 670 |
** ulimit to protect processes from each other, |
| 671 |
** ⎕SYL[1 2 3;2] to limit what the GNU APL interpreter may try to allocate, and |
| 672 |
** ⎕SYL[31 32;2] to avoid near-memory-starvation situations in the WS FULL |
| 673 |
exception handling. |
| 674 |
|
| 675 |
Thanks for using GNU APL. |
| 676 |
|