[ANNOUNCE] numactl 0.9 released

[ANNOUNCE] numactl 0.9 released

[Andi Kleen]

A new release of the numactl / libnuma package has been released

ftp://ftp.suse.com/pub/people/ak/numa/numactl-0.9.tar.gz
af9f10f1f65a88b3368c02157718aa58  numactl-0.9.tar.gz

It consists of a numactl program to run other programs with a specific 
NUMA policy and a libnuma shared library ("NUMA API") to set NUMA policy in 
applications and some additional tools.

The 0.8 release had some problems (in fact it didn't even build on !x86-64
without tweaks) so I did 0.9 early. I fixed all the known bugs, added
some patches that were reported after releases and added some minor new features.

The compile flags can be now changed with the standard CFLAGS=..
argument, no need anymore for OPT_CFLAGS (thanks Ian!)

I tweaked numademo a bit and it should be more useful now. It has more
stable measurements now and reports the the results in MB/s instead
of the weird units used before. There is a new random test that tests the 
memory performance with randomized accesses to defeat any hardware prefetching.

And ppc64 support should really work now (I hope at least) 

And numastat now knows how to wrap the display for a large number
of nodes (a killer feature!)

I hope to declare this numactl 1.0 soon unless any bad bugs are reported.

Any feedback appreciated.

-Andi

Detailed changes since 0.8:
- Get rid of bogus distance.o that broke compilation on !x86-64 (sorry)
- Handle CFLAGS overriding without OPT_CFLAGS (Ian Wienand)
- Fix up section of get/set_mempolicy (Ian Wienand)
- When no NUMA available fall back to one global node instead of one node
per CPU (Samuel Thibault)
- Don't rely on architecture symbols for dependency generation
- Use __powerpc__ to detect PPC/PPC64
- numastat: 
  * wrap display properly with many nodes
  * display nodes in forward order
  * install manpage in `make install'.
- remove bogus numamemcpy.c
- numademo: 
  * allow standalone compile, make streamlib optional
  * clean up output
  * change output unit to standard MB/s
  * compile with more optimization
  * add random pass to fool any prefetching (slow)
- make numademo compileable outside source tree
- use gettimeofday instead of time stamp counters in benchmarks
- support valgrind in testsuite
- other minor changes

Re: [ANNOUNCE] numactl 0.9 released

[Christoph Lameter]

Here is a patch in order to make numactl support page migration.

Comments welcome.

Index: numactl-0.9/mbind.2
=================================--- numactl-0.9.orig/mbind.2	2004-06-06 07:12:13.000000000 -0700
+++ numactl-0.9/mbind.2	2006-01-06 10:17:49.000000000 -0800
@@ -67,6 +67,30 @@ parameter 
 will be returned when the existing pages in the mapping don't follow
 the policy.
 
+When
+.B MPOL_MF_MOVE
+is passed in the 
+.B flags
+then attempts will be made to move all the pages in the mapping
+so that they follow the policy. Pages that are shared with other
+processes are not moved. If
+.B MPOL_MF_STRICT
+is also specified then
+.I EIO
+will be returned if some pages could not be moved.
+
+When
+.B MPOL_MF_MOVE_ALL
+is passed in the 
+.B flags
+then all pages in the mapping will be moved regardless of whether
+other processes use the pages. The process specifying this flag must
+have administrative priviledges. If
+.B MPOL_MF_STRICT
+is also specified then
+.I EIO
+will be returned if some pages could not be moved.
+
 The 
 .I MPOL_DEFAULT
 policy is the default and means to use the underlying process policy
@@ -133,6 +157,9 @@ header.
 is ignored on huge page mappings right now. For preferred and interleave 
 mappings it will only accept the first choice node.
 
+.I MPOL_MF_MOVE_*
+is only available on Linux 2.6.16 and later.
+
 For 
 .I MPOL_INTERLEAVE
 mode the interleaving is changed at fault time. The final layout of 
Index: numactl-0.9/numaif.h
=================================--- numactl-0.9.orig/numaif.h	2005-02-11 02:26:47.000000000 -0800
+++ numactl-0.9/numaif.h	2006-01-06 10:59:12.000000000 -0800
@@ -15,6 +15,8 @@ extern long mbind(void *start, unsigned 
 		  const unsigned long *nmask, unsigned long maxnode, unsigned flags);
 extern long set_mempolicy(int mode, const unsigned long *nmask, 
 			  unsigned long maxnode);
+extern long migratepages(int pid, unsigned long maxnode, unsigned long *fromnode,
+			unsigned long *tonode);
 
 /* Policies */
 #define MPOL_DEFAULT     0
@@ -30,6 +32,8 @@ extern long set_mempolicy(int mode, cons
 
 /* Flags for mbind */
 #define MPOL_MF_STRICT  (1<<0)  /* Verify existing pages in the mapping */
+#define MPOL_MF_MOVE	(1<<1)  /* Move pages owned by this process to conform to mapping */
+#define MPOL_MF_MOVE_ALL (1<<2) /* Move every page to conform to mapping */
 
 #ifdef __cplusplus
 }
Index: numactl-0.9/Makefile
=================================--- numactl-0.9.orig/Makefile	2006-01-03 12:36:13.000000000 -0800
+++ numactl-0.9/Makefile	2006-01-06 12:12:14.000000000 -0800
@@ -25,12 +25,14 @@ prefix := /usr
 libdir := ${prefix}$(shell if [ -d /usr/lib64 ] ; then echo "/lib64" ; else echo "/lib"  ; fi)
 docdir := ${prefix}/share/doc
 
-all: numactl libnuma.so numademo numamon memhog test/tshared stream \
+all: numactl migratepages libnuma.so numademo numamon memhog test/tshared stream \
      test/mynode test/pagesize test/ftok test/prefered test/randmap \
 	 test/nodemap test/distance
 
 numactl: numactl.o util.o shm.o bitops.o libnuma.so
 
+migratepages: migratepages.c util.o bitops.o libnuma.so
+
 util.o: util.c
 
 memhog: util.o memhog.o libnuma.so
@@ -94,10 +96,11 @@ set_membind set_preferred set_strict set
 tonodemask_memory distance
 
 MANPAGES := numa.3 numactl.8 mbind.2 set_mempolicy.2 get_mempolicy.2 \
-	    numastat.8
+	    numastat.8 migratepages.8
 
-install: numactl numademo.c numamon memhog libnuma.so.1 numa.h numaif.h numastat ${MANPAGES}
+install: numactl migratepages numademo.c numamon memhog libnuma.so.1 numa.h numaif.h numastat ${MANPAGES}
 	cp numactl ${prefix}/bin
+	cp migratepages ${prefix}/bin
 	cp numademo ${prefix}/bin
 	cp memhog ${prefix}/bin
 	cp set_mempolicy.2 ${prefix}/share/man/man2
Index: numactl-0.9/migratepages.c
=================================--- /dev/null	1970-01-01 00:00:00.000000000 +0000
+++ numactl-0.9/migratepages.c	2006-01-06 12:20:11.000000000 -0800
@@ -0,0 +1,97 @@
+/*
+ * Copyright (C) 2005 Christoph Lameter, Silicon Graphics, Incorporated.
+ * based on Andi Kleen's numactl.c.
+ *
+ * Manual process migration
+ *
+ * migratepages is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public
+ * License as published by the Free Software Foundation; version 2.
+ *
+ * migratepages is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should find a copy of v2 of the GNU General Public License somewhere
+ * on your Linux system; if not, write to the Free Software Foundation,
+ * Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
+ */
+
+#define _GNU_SOURCE
+#include <getopt.h>
+#include <errno.h>
+#include <stdio.h> 
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+#include <stdarg.h>
+#include "numaif.h"
+#include "numa.h"
+#include "numaint.h"
+#include "util.h"
+
+struct option opts[] = {
+	{"help", 0, 0, 'h' },
+	{ 0 }
+};
+
+void usage(void)
+{
+	fprintf(stderr,
+		"usage: migratepages pid from-nodes to-nodes\n"
+		"\n"
+		"nodes is a comma delimited list of node numbers or A-B ranges or none/all.\n"
+);
+	exit(1);
+}
+
+void checknuma(void)
+{
+	static int numa = -1;
+	if (numa < 0) {
+		if (numa_available() < 0)
+			complain("This system does not support NUMA functionality");
+	}
+	numa = 0;
+}
+
+int main(int argc, char *argv[])
+{
+	int c;
+	char *end;
+	int rc;
+	int pid;
+	nodemask_t fromnodes;
+	nodemask_t tonodes;
+
+	while ((c = getopt_long(argc,argv,"h", opts, NULL)) != -1) {
+		switch (c) {
+		default:
+			usage();
+		}
+	}
+
+	argv += optind;
+	argc -= optind;
+
+	if (argc != 4)
+		usage();
+
+	checknuma();
+
+	pid = strtoul(argv[1], &end, 0);
+	if (*end)
+		usage();
+
+	fromnodes = nodemask(argv[2]);
+	tonodes = nodemask(argv[3]);
+
+	rc = numa_migrate_pages(pid, &fromnodes, &tonodes);
+
+	if (rc) {
+		perror("migrate_pages");
+		return 1;
+	}
+	return 0;
+}
Index: numactl-0.9/syscall.c
=================================--- numactl-0.9.orig/syscall.c	2006-01-03 10:49:17.000000000 -0800
+++ numactl-0.9/syscall.c	2006-01-06 11:55:36.000000000 -0800
@@ -35,10 +35,12 @@
 #define __NR_mbind 237
 #define __NR_set_mempolicy 238
 #define __NR_get_mempolicy 239
+#define __NR_migrate_pages 256
 
 #elif defined(__ia64__)
 #define __NR_sched_setaffinity    1231
 #define __NR_sched_getaffinity    1232
+#define __NR_migrate_pages	1280
 
 /* Official allocation */
 
@@ -51,12 +53,14 @@
 #define __NR_mbind 274
 #define __NR_get_mempolicy 275
 #define __NR_set_mempolicy 276
+#define __NR_migrate_pages 294
 
 #elif defined(__powerpc__)
 
 #define __NR_mbind 259
 #define __NR_get_mempolicy 260
 #define __NR_set_mempolicy 261
+#define __NR_migrate_pages 280
 
 #elif !defined(DEPS_RUN)
 #error "Add syscalls for your architecture or update kernel headers"
@@ -141,6 +145,12 @@ long WEAK set_mempolicy(int mode, const 
 	return syscall(__NR_set_mempolicy,mode,nmask,maxnode);
 }
 
+long WEAK migrate_pages(int pid, unsigned long maxnode,
+	const unsigned long *frommask, const unsigned long *tomask)
+{
+	return syscall(__NR_migrate_pages, pid, maxnode, frommask, tomask);
+}
+
 /* SLES8 glibc doesn't define those */
 
 int numa_sched_setaffinity(pid_t pid, unsigned len, const unsigned long *mask)
@@ -159,3 +169,5 @@ make_internal_alias(numa_sched_setaffini
 make_internal_alias(get_mempolicy);
 make_internal_alias(set_mempolicy);
 make_internal_alias(mbind);
+make_internal_alias(migrate_pages);
+
Index: numactl-0.9/numa.h
=================================--- numactl-0.9.orig/numa.h	2005-12-25 14:20:34.000000000 -0800
+++ numactl-0.9/numa.h	2006-01-06 11:40:25.000000000 -0800
@@ -176,6 +176,8 @@ extern int numa_exit_on_error;
    once. */
 void numa_warn(int num, char *fmt, ...);
 
+int numa_migrate_pages(int pid, const nodemask_t *from, const nodemask_t *to);
+
 #ifdef __cplusplus
 }
 #endif
Index: numactl-0.9/libnuma.c
=================================--- numactl-0.9.orig/libnuma.c	2005-12-19 04:11:51.000000000 -0800
+++ numactl-0.9/libnuma.c	2006-01-06 12:00:03.000000000 -0800
@@ -600,6 +600,19 @@ nodemask_t numa_get_run_node_mask(void)
 	return mask;
 } 
 
+int numa_migrate_pages(int pid, const nodemask_t *fromnodes, const nodemask_t *tonodes)
+{
+	int err;
+
+	err = migrate_pages(pid, NUMA_NUM_NODES + 1, &fromnodes->n[0], &tonodes->n[0]);
+
+	if (err < 0) {
+		errno = -err;
+		return -1;
+	}
+	return err;
+}
+
 int numa_run_on_node(int node)
 { 
 	int ncpus = number_of_cpus();
Index: numactl-0.9/numaint.h
=================================--- numactl-0.9.orig/numaint.h	2005-04-28 04:40:38.000000000 -0700
+++ numactl-0.9/numaint.h	2006-01-06 11:56:36.000000000 -0800
@@ -11,7 +11,9 @@ extern long mbind_int(void *start, unsig
 		  const unsigned long *nmask, unsigned long maxnode, unsigned flags);
 extern long set_mempolicy_int(int mode, const unsigned long *nmask, 
 			  unsigned long maxnode);
-                                                    
+extern long migrate_pages(int pid, unsigned long maxnode, const unsigned long *frommask,
+	const unsigned long *tomask);
+
 #define SHM_HUGETLB     04000   /* segment will use huge TLB pages */
 
 #define CPU_BYTES(x) (round_up(x, BITS_PER_LONG)/8)
Index: numactl-0.9/migratepages.8
=================================--- /dev/null	1970-01-01 00:00:00.000000000 +0000
+++ numactl-0.9/migratepages.8	2006-01-06 12:18:03.000000000 -0800
@@ -0,0 +1,63 @@
+.\" t
+.\" Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
+.\"
+.\" based on Andi Kleen's numactl manpage
+.\"
+.TH MIGRATEPAGES 8 "Jan 2005" "SGI" "Linux Administrator's Manual"
+.SH NAME
+migratepages \- Migrate the physical location of pages of a process
+.SH SYNOPSIS
+.B migratepages
+pid from-nodes to-nodes
+.SH DESCRIPTION
+.B migratepages
+moves the physical localtion of a processes pages without any changes of the
+virtual address space of the process. This is usually done to optimize
+the performance of a process by moving the pages near to the processor
+executing a process.
+.TP
+Valid node specifiers
+.TS
+tab(:);
+l l. 
+all:All nodes
+number:Node number
+number1{,number2}:Node number1 and Node number2
+number1-number2:Nodes from number1 to number2
+! nodes:Invert selection of the following specification.
+.TE
+.SH NOTES
+Requires an NUMA policy aware kernel.
+
+migratepages will only move pages that are not shared with other
+processes if called by a user without administrative priviledges (but
+with the right to modify the process).
+
+migratepages will move all pages if invoked from root (or a user with
+administrative priviledges).
+
+.SH FILES
+.I /proc/<pid>/numastat
+for information about the NUMA memory use of a process.
+.SH COPYRIGHT
+Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
+migratepages is under the GNU General Public License, v.2
+
+.SH SEE ALSO
+.I numactl(8)
+,
+.I set_mempolicy(2)
+,
+.I get_mempolicy(2)
+,
+.I mbind(2)
+,
+.I sched_setaffinity(2)
+, 
+.I sched_getaffinity(2)
+,
+.I proc(5)
+, 
+.I ftok(3)
+,
+.I shmat(2)
Index: numactl-0.9/numactl.8
=================================--- numactl-0.9.orig/numactl.8	2005-12-16 04:13:19.000000000 -0800
+++ numactl-0.9/numactl.8	2006-01-06 12:11:31.000000000 -0800
@@ -271,3 +271,6 @@ numactl and the demo programs are under 
 .I ftok(3)
 ,
 .I shmat(2)
+,
+.I migratepages(8)
+

Re: [ANNOUNCE] numactl 0.9 released

[Andi Kleen]

On Friday 06 January 2006 21:24, Christoph Lameter wrote:
> Here is a patch in order to make numactl support page migration.

Thanks,

Did migrate_pages make it into Linus' tree yet? I don't see it there yet. 

I think the name of the command is a bit unintuitive - how 
about calling it changenodemem or movenodemem ? (better suggestions 
welcome) 

At least some words in the manpage what happens when you
specify multiple nodes in from/to might be a good idea.



> +	pid = strtoul(argv[1], &end, 0);
> +	if (*end)

|| end = argv[1]


> +moves the physical localtion of a processes pages without any changes of the

typo.


> +.SH COPYRIGHT
> +Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
> +migratepages is under the GNU General Public License, v.2

I can only warn you of putting your name visibly into the manpages - 
it will give you a constant stream of mails from confused users for many years ;-)

> +,
> +.I shmat(2)

Refer to taskset too? (or alternatively add a option to set the CPU
affinity using cpus or nodes too - taskset has the disadvantage that you cannot 
specify nodes so a user would need to know the cpu<->node mapping) 


-Andi

Re: [ANNOUNCE] numactl 0.9 released

[Christoph Lameter]

On Fri, 6 Jan 2006, Andi Kleen wrote:

> Did migrate_pages make it into Linus' tree yet? I don't see it there yet. 

Not yet but I would expect it for 2.6.16. There is also an open suse RFE 
for inclusion of migrate_pages into SLES10. 

> I think the name of the command is a bit unintuitive - how 
> about calling it changenodemem or movenodemem ? (better suggestions 
> welcome) 

The name is a bit unintuitive but its named the same way as the kernel 
function. It moves pages of a process so I guess the name is good enough 
unless someone comes up with a better name.

Here is a patch to fix the various issues you pointed out. Thank you:

Index: numactl-0.9/migratepages.8
=================================--- numactl-0.9.orig/migratepages.8	2006-01-06 12:18:03.000000000 -0800
+++ numactl-0.9/migratepages.8	2006-01-06 22:29:25.000000000 -0800
@@ -11,10 +11,16 @@ migratepages \- Migrate the physical loc
 pid from-nodes to-nodes
 .SH DESCRIPTION
 .B migratepages
-moves the physical localtion of a processes pages without any changes of the
+moves the physical location of a processes pages without any changes of the
 virtual address space of the process. This is usually done to optimize
 the performance of a process by moving the pages near to the processor
 executing a process.
+
+If multiple nodes are specified for from-nodes or to-nodes then
+an attempt is made to preserve the relative location of
+each page in each nodeset. F.e. if we move from nodes 2-5 to
+7,9,12-13 then the preferred mode of transfer is to move pages from
+2->7, 3->9, 4->12 and 5->13. 
 .TP
 Valid node specifiers
 .TS
@@ -61,3 +67,6 @@ migratepages is under the GNU General Pu
 .I ftok(3)
 ,
 .I shmat(2)
+,
+.I taskset(1)
+
Index: numactl-0.9/migratepages.c
=================================--- numactl-0.9.orig/migratepages.c	2006-01-06 22:29:44.000000000 -0800
+++ numactl-0.9/migratepages.c	2006-01-06 22:29:47.000000000 -0800
@@ -81,7 +81,7 @@ int main(int argc, char *argv[])
 	checknuma();
 
 	pid = strtoul(argv[1], &end, 0);
-	if (*end)
+	if (*end || end = argv[1])
 		usage();
 
 	fromnodes = nodemask(argv[2]);