Skip to contentSkip to conversation

Automatic NUMA placement for xl

PATCH Superseded Dario Faggioli 2 participants 5 messages Jul 4, 2012

lore
Share & tools
lore
#0 lore raw

PATCH v3 1/1 Automatic NUMA placement for xl

Hello,

Third version of the NUMA placement series Xen 4.2.

All the comments received during v2's review have been addressed (more details in single changelogs).

The most notable changes are the following:
- the libxl_cpumap --> libxl_bitmap renaming has been rebased on top of the recent patches that allows us to allocate bitmaps of different sizes;
- the heuristics for deciding which NUMA placement is the best one has been redesigned, so that it now provides total ordering.

Here it is what this posting contains (* = acked during previous round):

* [PATCH 01 of 10 v3] libxl: add a new Array type to the IDL [PATCH 02 of 10 v3] libxl,libxc: introduce libxl_get_numainfo()
* [PATCH 03 of 10 v3] xl: add more NUMA information to `xl info -n' [PATCH 04 of 10 v3] libxl: rename libxl_cpumap to libxl_bitmap [PATCH 05 of 10 v3] libxl: expand the libxl_bitmap API a bit
* [PATCH 06 of 10 v3] libxl: introduce some node map helpers [PATCH 07 of 10 v3] libxl: explicitly check for libmath in autoconf

Is where data structures, utility functions and infrastructure are introduced.

* [PATCH 08 of 10 v3] libxl: enable automatic placement of guests on NUMA nodes
* [PATCH 09 of 10 v3] libxl: have NUMA placement deal with cpupools

Host the core of the mechanism.

* [PATCH 10 of 10 v3] Some automatic NUMA placement documentation

For some more documentation.

Thanks a lot and Regards,
Dario

Signature · 4 lines
<<This happens because I choose it to happen!>> (Raistlin Majere)
-----------------------------------------------------------------
Dario Faggioli, Ph.D, http://retis.sssup.it/people/faggioli
Senior Software Engineer, Citrix Systems R&D Ltd., Cambridge (UK)
#1 lore raw

PATCH v3 1/1 libxl: add a new Array type to the IDL

# HG changeset patch
# User Ian Campbell <[email protected]> # Date 1341416322 -7200
# Node ID 8e367818e194c212cd1470aad663f3243ff53bdb # Parent 42f76d536b116d2ebad1b6705ae51ecd171d2581
libxl: add a new Array type to the IDL

And make all the required infrastructure updates to enable this.

Since there are currently no uses of this type there is no change to the generated code.

---
Changes from v2:
* v2's patch replaced by what Ian Campbell posted as <[email protected]>, as agreed during the review process of that patch.

tools/libxl/gentest.py +12 -0
@@ -27,6 +27,18 @@ def gen_rand_init(ty, v, indent = "    "
     s = ""
     if isinstance(ty, idl.Enumeration):
         s += "%s = %s;\n" % (ty.pass_arg(v, parent is None), randomize_enum(ty))
+    elif isinstance(ty, idl.Array):
+        if parent is None:
+            raise Exception("Array type must have a parent")
+        s += "%s = rand()%%8;\n" % (parent + ty.lenvar.name)
+        s += "%s = calloc(%s, sizeof(*%s));\n" % \
+            (v, parent + ty.lenvar.name, v)
+        s += "{\n"
+        s += "    int i;\n"
+        s += "    for (i=0; i<%s; i++)\n" % (parent + ty.lenvar.name)
+        s += gen_rand_init(ty.elem_type, v+"[i]",
+                           indent + "        ", parent)
+        s += "}\n"
     elif isinstance(ty, idl.KeyedUnion):
         if parent is None:
             raise Exception("KeyedUnion type must have a parent")
tools/libxl/gentypes.py +39 -3

60 lines not on this page · raw message

tools/libxl/idl.py +11 -0
@@ -266,6 +266,17 @@ string = Builtin("char *", namespace = N
                  json_fn = "libxl__string_gen_json",
                  autogenerate_json = False)
 
+class Array(Type):
+    """An array of the same type"""
+    def __init__(self, elem_type, lenvar_name, **kwargs):
+        kwargs.setdefault('dispose_fn', 'free')
+        Type.__init__(self, namespace=elem_type.namespace, typename=elem_type.rawname + " *", **kwargs)
+
+        lv_kwargs = dict([(x.lstrip('lenvar_'),y) for (x,y) in kwargs.items() if x.startswith('lenvar_')])
+
+        self.lenvar = Field(integer, lenvar_name, **lv_kwargs)
+        self.elem_type = elem_type
+
 class OrderedDict(dict):
     """A dictionary which remembers insertion order.
 
tools/libxl/idl.txt +17 -4
@@ -145,11 +145,24 @@ idl.KeyedUnion
 
  A subclass of idl.Aggregate which represents the C union type
  where the currently valid member of the union can be determined based
- upon another member in the containing type.
+ upon another member in the containing type. An idl.KeyedUnion must
+ always be a member of a containing idl.Aggregate type.
 
- The KeyedUnion.keyvar contains an idl.type the member of the
- containing type which determines the valid member of the union. The
- must be an instance of the Enumeration type.
+ The KeyedUnion.keyvar contains an idl.Field, this is the member of
+ the containing type which determines the valid member of the
+ union. The idl.Field.type of the keyvar must be an Enumeration type.
+
+idl.Array
+
+  A class representing an array of similar elements. An idl.Array must
+  always be an idl.Field of a containing idl.Aggregate.
+
+  idl.Array.elem_type contains an idl.Type which is the type of each
+  element of the array.
+
+  idl.Array.len_var contains an idl.Field which is added to the parent
+  idl.Aggregate and will contain the length of the array. The field
+  MUST be named num_ARRAYNAME.
 
 Standard Types
 --------------
tools/ocaml/libs/xl/genwrap.py +14 -1
@@ -55,7 +55,8 @@ def ocaml_type_of(ty):
             return "int%d" % ty.width
         else:
             return "int"
-
+    elif isinstance(ty,idl.Array):
+        return "%s array" % ocaml_type_of(ty.elem_type)
     elif isinstance(ty,idl.Builtin):
         if not builtins.has_key(ty.typename):
             raise NotImplementedError("Unknown Builtin %s (%s)" % (ty.typename, type(ty)))
@@ -138,6 +139,8 @@ def c_val(ty, c, o, indent="", parent = 
         if not fn:
             raise NotImplementedError("No c_val fn for Builtin %s (%s)" % (ty.typename, type(ty)))
         s += "%s;" % (fn % { "o": o, "c": c })
+    elif isinstance (ty,idl.Array):
+        raise("Cannot handle Array type\n")
     elif isinstance(ty,idl.Enumeration) and (parent is None):
         n = 0
         s += "switch(Int_val(%s)) {\n" % o
@@ -195,6 +198,16 @@ def ocaml_Val(ty, o, c, indent="", paren
         if not fn:
             raise NotImplementedError("No ocaml Val fn for Builtin %s (%s)" % (ty.typename, type(ty)))
         s += "%s = %s;" % (o, fn % { "c": c })
+    elif isinstance(ty, idl.Array):
+        s += "{\n"
+        s += "\t    int i;\n"
+        s += "\t    value array_elem;\n"
+        s += "\t    %s = caml_alloc(%s,0);\n" % (o, parent + ty.lenvar.name)
+        s += "\t    for(i=0; i<%s; i++) {\n" % (parent + ty.lenvar.name)
+        s += "\t        %s\n" % ocaml_Val(ty.elem_type, "array_elem", c + "[i]", "")
+        s += "\t        Store_field(%s, i, array_elem);\n" % o
+        s += "\t    }\n"
+        s += "\t}"
     elif isinstance(ty,idl.Enumeration) and (parent is None):
         n = 0
         s += "switch(%s) {\n" % c
tools/python/genwrap.py +5 -3
@@ -4,7 +4,7 @@ import sys,os
 
 import idl
 
-(TYPE_DEFBOOL, TYPE_BOOL, TYPE_INT, TYPE_UINT, TYPE_STRING, TYPE_AGGREGATE) = range(6)
+(TYPE_DEFBOOL, TYPE_BOOL, TYPE_INT, TYPE_UINT, TYPE_STRING, TYPE_ARRAY, TYPE_AGGREGATE) = range(7)
 
 def py_type(ty):
     if ty == idl.bool:
@@ -18,6 +18,8 @@ def py_type(ty):
             return TYPE_INT
         else:
             return TYPE_UINT
+    if isinstance(ty, idl.Array):
+        return TYPE_ARRAY
     if isinstance(ty, idl.Aggregate):
         return TYPE_AGGREGATE
     if ty == idl.string:
@@ -74,7 +76,7 @@ def py_attrib_get(ty, f):
         l.append('    return genwrap__ull_get(self->obj.%s);'%f.name)
     elif t == TYPE_STRING:
         l.append('    return genwrap__string_get(&self->obj.%s);'%f.name)
-    elif t == TYPE_AGGREGATE:
+    elif t == TYPE_AGGREGATE or t == TYPE_ARRAY:
         l.append('    PyErr_SetString(PyExc_NotImplementedError, "Getting %s");'%ty.typename)
         l.append('    return NULL;')
     else:
@@ -105,7 +107,7 @@ def py_attrib_set(ty, f):
         l.append('    return ret;')
     elif t == TYPE_STRING:
         l.append('    return genwrap__string_set(v, &self->obj.%s);'%f.name)
-    elif t == TYPE_AGGREGATE:
+    elif t == TYPE_AGGREGATE or t == TYPE_ARRAY:
         l.append('    PyErr_SetString(PyExc_NotImplementedError, "Setting %s");'%ty.typename)
         l.append('    return -1;')
     else:
#2 lore raw

PATCH v3 1/1 libxl, libxc: introduce libxl_get_numainfo()

In reply to Dario Faggioli

# HG changeset patch
# User Dario Faggioli <[email protected]> # Date 1341416323 -7200
# Node ID 0ca91a203fc95d3d18bb436ecdc7106b0b2ff22f # Parent 8e367818e194c212cd1470aad663f3243ff53bdb libxl,libxc: introduce libxl_get_numainfo()

Make some NUMA node information available to the toolstack. Achieve this by means of xc_numainfo(), which exposes memory size and amount of free memory of each node, as well as the relative distances of each node to all the others.

For properly exposing distances we need the IDL to support arrays.

---
Changes from v2:
* converted libxl__zalloc(NULL, ...) to libxl_calloc(NOGC, ...).
* Fixed the comment about memory ownership of libxl_get_numainfo().
* Added a comment for libxl_numainfo in libxl_types.idl.

Changes from v1:
* malloc converted to libxl__zalloc(NOGC, ...).
* The patch also accommodates some bits of what was in "libxc,
libxl: introduce xc_nodemap_t and libxl_nodemap" which was removed as well, as full support for node maps at libxc level is not needed (yet!).

tools/libxc/xc_misc.c +14 -0
@@ -35,6 +35,20 @@ int xc_get_max_cpus(xc_interface *xch)
     return max_cpus;
 }
 
+int xc_get_max_nodes(xc_interface *xch)
+{
+    static int max_nodes = 0;
+    xc_physinfo_t physinfo;
+
+    if ( max_nodes )
+        return max_nodes;
+
+    if ( !xc_physinfo(xch, &physinfo) )
+        max_nodes = physinfo.max_node_id + 1;
+
+    return max_nodes;
+}
+
 int xc_get_cpumap_size(xc_interface *xch)
 {
     return (xc_get_max_cpus(xch) + 7) / 8;
tools/libxc/xenctrl.h +6 -0
@@ -329,6 +329,12 @@ int xc_get_cpumap_size(xc_interface *xch
 /* allocate a cpumap */
 xc_cpumap_t xc_cpumap_alloc(xc_interface *xch);
 
+ /*
+ * NODEMAP handling
+ */
+/* return maximum number of NUMA nodes the hypervisor supports */
+int xc_get_max_nodes(xc_interface *xch);
+
 /*
  * DOMAIN DEBUGGING FUNCTIONS
  */
tools/libxl/libxl.c +69 -0

75 lines not on this page · raw message

tools/libxl/libxl.h +7 -0
@@ -532,6 +532,9 @@ int libxl_domain_preserve(libxl_ctx *ctx
 /* get max. number of cpus supported by hypervisor */
 int libxl_get_max_cpus(libxl_ctx *ctx);
 
+/* get max. number of NUMA nodes supported by hypervisor */
+int libxl_get_max_nodes(libxl_ctx *ctx);
+
 int libxl_domain_rename(libxl_ctx *ctx, uint32_t domid,
                         const char *old_name, const char *new_name);
 
@@ -604,6 +607,10 @@ void libxl_vminfo_list_free(libxl_vminfo
 libxl_cputopology *libxl_get_cpu_topology(libxl_ctx *ctx, int *nb_cpu_out);
 void libxl_cputopology_list_free(libxl_cputopology *, int nb_cpu);
 
+#define LIBXL_NUMAINFO_INVALID_ENTRY (~(uint32_t)0)
+libxl_numainfo *libxl_get_numainfo(libxl_ctx *ctx, int *nr);
+void libxl_numainfo_list_free(libxl_numainfo *, int nr);
+
 libxl_vcpuinfo *libxl_list_vcpu(libxl_ctx *ctx, uint32_t domid,
                                 int *nb_vcpu, int *nr_vcpus_out);
 void libxl_vcpuinfo_list_free(libxl_vcpuinfo *, int nr_vcpus);
tools/libxl/libxl_types.idl +9 -0
@@ -433,6 +433,15 @@ libxl_physinfo = Struct("physinfo", [
     ("cap_hvm_directio", bool),
     ], dir=DIR_OUT)
 
+# NUMA node characteristics: size and free are how much memory it has, and how
+# much of it is free, respectively. dists is an array of distances from this
+# node to each other node.
+libxl_numainfo = Struct("numainfo", [
+    ("size", uint64),
+    ("free", uint64),
+    ("dists", Array(uint32, "num_dists")),
+    ], dir=DIR_OUT)
+
 libxl_cputopology = Struct("cputopology", [
     ("core", uint32),
     ("socket", uint32),
tools/libxl/libxl_utils.c +13 -0
@@ -572,6 +572,11 @@ int libxl_get_max_cpus(libxl_ctx *ctx)
     return xc_get_max_cpus(ctx->xch);
 }
 
+int libxl_get_max_nodes(libxl_ctx *ctx)
+{
+    return xc_get_max_nodes(ctx->xch);
+}
+
 int libxl__enum_from_string(const libxl_enum_string_table *t,
                             const char *s, int *e)
 {
@@ -594,6 +599,14 @@ void libxl_cputopology_list_free(libxl_c
     free(list);
 }
 
+void libxl_numainfo_list_free(libxl_numainfo *list, int nr)
+{
+    int i;
+    for (i = 0; i < nr; i++)
+        libxl_numainfo_dispose(&list[i]);
+    free(list);
+}
+
 void libxl_vcpuinfo_list_free(libxl_vcpuinfo *list, int nr)
 {
     int i;
xen/include/public/sysctl.h +1 -0
@@ -484,6 +484,7 @@ typedef struct xen_sysctl_topologyinfo x
 DEFINE_XEN_GUEST_HANDLE(xen_sysctl_topologyinfo_t);
 
 /* XEN_SYSCTL_numainfo */
+#define INVALID_NUMAINFO_ID (~0U)
 struct xen_sysctl_numainfo {
     /*
      * IN: maximum addressable entry in the caller-provided arrays.
#3 lore raw

PATCH v3 1/1 xl: add more NUMA information to `xl info -n'

In reply to Dario Faggioli

# HG changeset patch
# User Dario Faggioli <[email protected]> # Date 1341416323 -7200
# Node ID f1227d5a82e56d10e302aec4c3717d281718a349 # Parent 0ca91a203fc95d3d18bb436ecdc7106b0b2ff22f
xl: add more NUMA information to `xl info -n'

So that the user knows how much memory there is on each node and how far they are from each others.

---
Changes from v1:
* integer division replaced by right shift.

tools/libxl/xl_cmdimpl.c +33 -3
@@ -4249,6 +4249,36 @@ static void output_physinfo(void)
     return;
 }
 
+static void output_numainfo(void)
+{
+    libxl_numainfo *info;
+    int i, j, nr;
+
+    info = libxl_get_numainfo(ctx, &nr);
+    if (info == NULL) {
+        fprintf(stderr, "libxl_get_numainfo failed.\n");
+        return;
+    }
+
+    printf("numa_info              :\n");
+    printf("node:    memsize    memfree    distances\n");
+
+    for (i = 0; i < nr; i++) {
+        if (info[i].size != LIBXL_NUMAINFO_INVALID_ENTRY) {
+            printf("%4d:    %6"PRIu64"     %6"PRIu64"      %d", i,
+                   info[i].size >> 20, info[i].free >> 20,
+                   info[i].dists[0]);
+            for (j = 1; j < info[i].num_dists; j++)
+                printf(",%d", info[i].dists[j]);
+            printf("\n");
+        }
+    }
+
+    libxl_numainfo_list_free(info, nr);
+
+    return;
+}
+
 static void output_topologyinfo(void)
 {
     libxl_cputopology *info;
@@ -4271,8 +4301,6 @@ static void output_topologyinfo(void)
 
     libxl_cputopology_list_free(info, nr);
 
-    printf("numa_info              : none\n");
-
     return;
 }
 
@@ -4282,8 +4310,10 @@ static void info(int numa)
 
     output_physinfo();
 
-    if (numa)
+    if (numa) {
         output_topologyinfo();
+        output_numainfo();
+    }
 
     output_xeninfo();
#4 lore raw
In reply to Dario Faggioli
Quoted 2 lines · Dario Faggioli

Hello,

Sorry for this! I don't have the slightest idea of why it `hg email' hanged after sending only these 3 mails. :-(

Anyway, my second attempt seems to have succeeded, so please, consider that for the full series (i.e., the thread where the msg-id of the introductory mail is <patchbomb.1341418679@Solace>), and sorry again for the noise.

Thanks and Regards,
Dario

Signature · 4 lines
<<This happens because I choose it to happen!>> (Raistlin Majere)
-----------------------------------------------------------------
Dario Faggioli, Ph.D, http://retis.sssup.it/people/faggioli
Senior Software Engineer, Citrix Systems R&D Ltd., Cambridge (UK)