[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[Xen-devel] [Patch] fix automatic NUMA placement



Hi,

the automatic NUMA placement code (find_relaxed_node) in XendDomainInfo.py contains several bugs:
1. The domain just created shall not be considered in the load calculation.
2. Nodes with too less memory for the domain must be considered during the load calculation phase and abandoned later, currently these nodes win with a load of 0.
3. CPUs which are offline shall not be considered in load calculation.
4. (Integer-)Dividing the load value by the number of cores per node may conceal subtle load differences resulting in a non-optimal node to be chosen.

Attached patch fixes these bugs.

Signed-off-by: Andre Przywara <andre.przywara@xxxxxxx>

Regards,
Andre.

P.S. If you don't like the lambda expression, you can use the more C-like version in the next line:
    if dom.domid == self.domid: continue

--
Andre Przywara
AMD-Operating System Research Center (OSRC), Dresden, Germany
Tel: +49 351 277-84917
----to satisfy European Law for business letters:
AMD Saxony Limited Liability Company & Co. KG,
Wilschdorfer Landstr. 101, 01109 Dresden, Germany
Register Court Dresden: HRA 4896, General Partner authorized
to represent: AMD Saxony LLC (Wilmington, Delaware, US)
General Manager of AMD Saxony LLC: Dr. Hans-R. Deppe, Thomas McCoy
diff -r e48453f82d30 tools/python/xen/xend/XendDomainInfo.py
--- a/tools/python/xen/xend/XendDomainInfo.py   Thu May 22 10:41:49 2008 +0100
+++ b/tools/python/xen/xend/XendDomainInfo.py   Thu May 22 13:48:36 2008 +0200
@@ -2091,28 +2091,28 @@ class XendDomainInfo:
                         xc.vcpu_setaffinity(self.domid, v, 
self.info['cpus'][v])
             else:
                 def find_relaxed_node(node_list):
-                    import sys 
+                    import sys
+                    nr_nodes = info['nr_nodes']
                     if node_list is None:
-                        node_list = range(0, info['nr_nodes'])
+                        node_list = range(0, nr_nodes)
                     nodeload = [0]
-                    nodeload = nodeload * info['nr_nodes']
+                    nodeload = nodeload * nr_nodes
                     from xen.xend import XendDomain
                     doms = XendDomain.instance().list('all')
-                    for dom in doms:
+                    for dom in filter (lambda d: d.domid != self.domid, doms):
                         cpuinfo = dom.getVCPUInfo()
                         for vcpu in sxp.children(cpuinfo, 'vcpu'):
-                            def vinfo(n, t):
-                                return t(sxp.child_value(vcpu, n))
-                            cpumap = vinfo('cpumap', list)
-                            for i in node_list:
+                            if sxp.child_value(vcpu, 'online') == 0: continue
+                            cpumap = list(sxp.child_value(vcpu,'cpumap'))
+                            for i in range(0, nr_nodes):
                                 node_cpumask = info['node_to_cpu'][i]
                                 for j in node_cpumask:
                                     if j in cpumap:
                                         nodeload[i] += 1
                                         break
-                    for i in node_list:
-                        if len(info['node_to_cpu'][i]) > 0:
-                            nodeload[i] = int(nodeload[i] / 
len(info['node_to_cpu'][i]))
+                    for i in range(0, nr_nodes):
+                        if len(info['node_to_cpu'][i]) > 0 and i in node_list:
+                            nodeload[i] = int(nodeload[i] * 16 / 
len(info['node_to_cpu'][i]))
                         else:
                             nodeload[i] = sys.maxint
                     index = nodeload.index( min(nodeload) )    
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel

 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.