[Pkg-ganeti-devel] [ganeti] 80/165: Add a design document on improving allocation

[Apollon Oikonomopoulos]

This is an automated email from the git hooks/post-receive script.

apoikos pushed a commit to branch master
in repository ganeti.

commit b274708c38e5d172f07f8158b32fd8b8686a8460
Author: Klaus Aehlig <aehlig at google.com>
Date:   Tue Mar 17 10:00:00 2015 +0100

    Add a design document on improving allocation
    
    This design describes an addition to the cluster
    metrics that improves allocation efficiency. So
    far, only the balancedness of the amount of reserved
    memory was taken into account, but not its total
    amount. The reserved memory, however, is not preserved
    by instance moves; it can be lowered by spreading the
    secondaries more widely over the cluster.
    
    Signed-off-by: Klaus Aehlig <aehlig at google.com>
    Reviewed-by: Petr Pudlak <pudlak at google.com>
---
 Makefile.am                          |  1 +
 doc/design-allocation-efficiency.rst | 59 ++++++++++++++++++++++++++++++++++++
 doc/design-draft.rst                 |  1 +
 3 files changed, 61 insertions(+)

diff --git a/Makefile.am b/Makefile.am
index 4a7a60b..5a8c300 100644
--- a/Makefile.am
+++ b/Makefile.am
@@ -653,6 +653,7 @@ docinput = \
 	doc/design-2.12.rst \
 	doc/design-2.13.rst \
 	doc/design-2.14.rst \
+	doc/design-allocation-efficiency.rst \
 	doc/design-autorepair.rst \
 	doc/design-bulk-create.rst \
 	doc/design-ceph-ganeti-support.rst \
diff --git a/doc/design-allocation-efficiency.rst b/doc/design-allocation-efficiency.rst
new file mode 100644
index 0000000..f375b27
--- /dev/null
+++ b/doc/design-allocation-efficiency.rst
@@ -0,0 +1,59 @@
+=========================================================================
+Improving allocation efficiency by considering the total reserved memory
+=========================================================================
+
+This document describes a change to the cluster metric to enhance
+the allocation efficiency of Ganeti's ``htools``.
+
+.. contents:: :depth: 4
+
+
+Current state and shortcomings
+==============================
+
+Ganeti's ``htools``, which typically make all allocation and balancing
+decisions, greedily try to improve the cluster metric. So it is important
+that the cluster metric faithfully reflects the objectives of these operations.
+Currently the cluster metric is composed of counting violations (instances on
+offline nodes, nodes that are not N+1 redundant, etc) and the sum of standard
+deviations of relative resource usage of the individual nodes. The latter
+component is to ensure that all nodes equally bear the load of the instances.
+This is reasonable for resources where the total usage is independent of
+its distribution, as it is the case for CPU, disk, and total RAM. It is,
+however, not true for reserved memory. By distributing its secondaries
+more widespread over the cluster, a node can reduce its reserved memory
+without increasing it on other nodes. Not taking this aspect into account
+has lead to quite inefficient allocation of instances on the cluster (see
+example below).
+
+Proposed changes
+================
+
+A new additive component is added to the cluster metric. It is the sum over
+all nodes of the fraction of reserved memory. This way, moves and allocations
+that reduce the amount of memory reserved to ensure N+1 redundancy are favored.
+
+Note that this component does not have the scaling of standard deviations of
+fractions, but, instead counts nodes reserved for N+1 redundancy. In an ideal
+allocation, this will not exceed 1. But bad allocations will violate this
+property. As waste of reserved memory is a more future-oriented problem than,
+e.g., current N+1 violations, we give the new component a relatively small
+weight of 0.25, so that counting current violations still dominate.
+
+Another consequence of this metric change is that the value 0 is no longer
+obtainable: as soon as we have DRBD instance, we have to reserve memory.
+However, in most cases only differences of scores influence decissions made.
+In the few cases, were absolute values of the cluster score are specified,
+they are interpreted as relative to the theoretical minimum of the reserved
+memory score.
+
+
+Example
+=======
+
+Consider the capacity of an empty cluster of 6 nodes, each capable of holding
+10 instances; this can be measured, e.g., by
+``hspace --simulate=p,6,204801,10241,21 --disk-template=drbd
+--standard-alloc=10240,1024,2``. Without the metric change 34 standard
+instances are allocated. With the metric change, 48 standard instances
+are allocated. This is a 41% increase in utilization.
diff --git a/doc/design-draft.rst b/doc/design-draft.rst
index 0f191cb..88f16ec 100644
--- a/doc/design-draft.rst
+++ b/doc/design-draft.rst
@@ -26,6 +26,7 @@ Design document drafts
    design-configlock.rst
    design-multi-storage-htools.rst
    design-dedicated-allocation.rst
+   design-allocation-efficiency.rst
 
 .. vim: set textwidth=72 :
 .. Local Variables:

-- 
Alioth's /usr/local/bin/git-commit-notice on /srv/git.debian.org/git/pkg-ganeti/ganeti.git