Re: mmap file performance
On Monday, 15 April 2024 at 16:13:41 UTC, Andy Valencia wrote: On Monday, 15 April 2024 at 08:05:25 UTC, Patrick Schluter wrote: The setup of a memory mapped file is relatively costly. For smaller files it is a net loss and read/write beats it hands down. Interestingly, this performance deficit is present even when run against the largest conveniently available file on my system--libQt6WebEngineCore.so.6.4.2 at 148 megs. But since this reproduces in its C counterpart, it is not at all a reflection of D. As you say, truly random access might play to mmap's strengths. Indeed, my statement concerning file size is misleading. It's the amount of operations done on the file that is important. For bigger files it is normal to have more operations. I have measurement for our system (Linux servers) where we have big index files which represent a ternary tree that are generally memory mapped. These files are several hundreds of megabytes big and the access is almost random. These files still grow but the growing parts are not memory mapped but accessed with pread() and pwrite() calls. Accesses via pread() take exactly twice the time from memory copy for reads of exactly 64 bytes (size of the record). My real point is that, whichever API I use, coding in D was far less tedious; I like the resulting code, and it showed no meaningful performance cost.
Re: mmap file performance
On Thursday, 11 April 2024 at 16:23:44 UTC, Andy Valencia wrote: [...] void main(in string[] argv) ^^ What if you want to use bool memorymapped; getopt (argv, "m", ); inside main? [1] Have you tried using "rm" [2] instead of "r" as stdioOpenmode under Linux for a "no code" method of employing mmap for reading? [1] const main args? [2011] https://forum.dlang.org/post/mailman.2277.1313189911.14074.digitalmars-d-le...@puremagic.com [2] section "NOTES" in fopen(3) https://man7.org/linux/man-pages/man3/fopen.3.html
Re: mmap file performance
On Monday, 15 April 2024 at 08:05:25 UTC, Patrick Schluter wrote: The setup of a memory mapped file is relatively costly. For smaller files it is a net loss and read/write beats it hands down. Interestingly, this performance deficit is present even when run against the largest conveniently available file on my system--libQt6WebEngineCore.so.6.4.2 at 148 megs. But since this reproduces in its C counterpart, it is not at all a reflection of D. As you say, truly random access might play to mmap's strengths. My real point is that, whichever API I use, coding in D was far less tedious; I like the resulting code, and it showed no meaningful performance cost.
Re: mmap file performance
On Thursday, 11 April 2024 at 00:24:44 UTC, Andy Valencia wrote: I wrote a "count newlines" based on mapped files. It used about twice the CPU of the version which just read 1 meg at a time. I thought something was amiss (needless slice indirection or something), so I wrote the code in C. It had the same CPU usage as the D version. So...mapped files, not so much. Not D's fault. And writing it in C made me realize how much easier it is to code in D! [...] The setup of a memory mapped file is relatively costly. For smaller files it is a net loss and read/write beats it hands down. Furthermore, sequential access is not the best way to exploit the advantages of mmap. Full random access is the strong suite of mmap as it replaces kernel syscalls (lseek,read, write or pread, pwrite) by user land processing. You could try MAP_POPULATE option in the mmap as it enables read-ahead on the file which may help on sequential code.
Re: mmap file performance
On Thursday, 11 April 2024 at 14:54:36 UTC, Steven Schveighoffer
wrote:
For a repeatable comparison, you should provide the code which
does 1MB reads.
With pleasure:
import std.stdio : writeln, File, stderr;
const uint BUFSIZE = 1024*1024;
private uint
countnl(File f)
{
uint res = 0;
char[BUFSIZE] buf;
while (!f.eof) {
auto sl = f.rawRead(buf);
foreach (c; sl) {
if (c == '\n') {
res += 1;
}
}
}
return res;
}
private uint
procfile(in string fn) {
import std.exception : ErrnoException;
File f;
try {
f = File(fn, "r");
} catch(ErrnoException e) {
stderr.writeln("Can't open: ", fn);
return 0;
}
uint res = countnl(f);
f.close();
return res;
}
void
main(in string[] argv)
{
foreach (fn; argv[1 .. $]) {
uint res;
res = procfile(fn);
writeln(fn, ": ", res);
}
}
Re: mmap file performance
On Thursday, 11 April 2024 at 00:24:44 UTC, Andy Valencia wrote: I wrote a "count newlines" based on mapped files. It used about twice the CPU of the version which just read 1 meg at a time. I thought something was amiss (needless slice indirection or something), so I wrote the code in C. It had the same CPU usage as the D version. So...mapped files, not so much. Not D's fault. And writing it in C made me realize how much easier it is to code in D! For a repeatable comparison, you should provide the code which does 1MB reads. I have found that mmapped files are faster than reading buffered files, but maybe only for large files? -Steve
mmap file performance
I wrote a "count newlines" based on mapped files. It used about
twice the CPU of the version which just read 1 meg at a time. I
thought something was amiss (needless slice indirection or
something), so I wrote the code in C. It had the same CPU usage
as the D version. So...mapped files, not so much. Not D's
fault. And writing it in C made me realize how much easier it is
to code in D!
The D version:
import std.stdio : writeln;
import std.mmfile : MmFile;
const uint CHUNKSZ = 65536;
size_t
countnl(ref shared char[] data)
{
size_t res = 0;
foreach (c; data) {
if (c == '\n') {
res += 1;
}
}
return res;
}
void
usage(in string progname)
{
import core.stdc.stdlib : exit;
import std.stdio : stderr;
stderr.writeln("Usage is: ", progname, " %s ...");
exit(1);
}
public:
void
main(string[] argv)
{
if (argv.length < 2) {
usage(argv[0]);
}
foreach(mn; argv[1 .. $]) {
auto mf = new MmFile(mn);
auto data = cast(shared char[])mf.opSlice();
size_t res;
res = countnl(data);
writeln(mn, ": ", res);
}
}
And the C one (no performance gain over D):
#include
#include
#include
#include
#include
static unsigned long
countnl(int fd, char *nm)
{
char *buf, *p;
struct stat st;
unsigned int cnt;
unsigned long res;
if (fstat(fd, ) < 0) {
perror(nm);
return(0);
}
cnt = st.st_size;
buf = mmap(0, cnt, PROT_READ, MAP_SHARED, fd, 0);
if (buf == MAP_FAILED) {
perror(nm);
return(0);
}
res = 0L;
for (p = buf; cnt; cnt -= 1) {
if (*p++ == '\n') {
res += 1L;
}
}
munmap(buf, st.st_size);
return(res);
}
int
main(int argc, char **argv)
{
int x;
for (x = 1; x < argc; ++x) {
unsigned long res;
char *nm = argv[x];
int fd = open(nm, O_RDONLY);
if (fd < 0) {
perror(nm);
continue;
}
res = countnl(fd, nm);
close(fd);
printf("%s: %uld\n", nm, res);
}
}
