|
1 |
===============================
|
|
2 |
Huge Pages Support for Ganeti
|
|
3 |
===============================
|
|
4 |
This is a design document about implementing support for huge pages in
|
|
5 |
Ganeti. (Please note that Ganeti works with Transparent Huge Pages i.e.
|
|
6 |
THP and any reference in this document to Huge Pages refers to explicit
|
|
7 |
Huge Pages).
|
|
8 |
|
|
9 |
Current State and Shortcomings:
|
|
10 |
-------------------------------
|
|
11 |
The Linux kernel allows using pages of larger size by setting aside a
|
|
12 |
portion of the memory. Using larger page size may enhance the
|
|
13 |
performance of applications that require a lot of memory by improving
|
|
14 |
page hits. To use huge pages, memory has to be reserved beforehand. This
|
|
15 |
portion of memory is subtracted from free memory and is considered as in
|
|
16 |
use. Currently Ganeti cannot take proper advantage of huge pages. On a
|
|
17 |
node, if huge pages are reserved and are available to fulfill the VM
|
|
18 |
request, Ganeti fails to recognize huge pages and considers the memory
|
|
19 |
reserved for huge pages as used memory. This leads to failure of
|
|
20 |
launching VMs on a node where memory is available in the form of huge
|
|
21 |
pages rather than normal pages.
|
|
22 |
|
|
23 |
Proposed Changes:
|
|
24 |
-----------------
|
|
25 |
The following components will be changed in order for Ganeti to take
|
|
26 |
advantage of Huge Pages.
|
|
27 |
|
|
28 |
Hypervisor Parameters:
|
|
29 |
----------------------
|
|
30 |
Currently, It is possible to set or modify huge pages mount point at
|
|
31 |
cluster level via the hypervisor parameter ``mem_path`` as::
|
|
32 |
|
|
33 |
$ gnt-cluster init --no-lvm-storage --no-drbd-storage \
|
|
34 |
>--enabled-hypervisors=kvm -nic-parameters link=br100 \
|
|
35 |
> -H kvm:mem_path=/mount/point/for/hugepages
|
|
36 |
|
|
37 |
This hypervisor parameter is inherited by all the instances as
|
|
38 |
default although it can be overriden at the instance level.
|
|
39 |
|
|
40 |
The following changes will be made to the inheritence behaviour.
|
|
41 |
|
|
42 |
- The hypervisor parameter ``mem_path`` and all other hypervisor
|
|
43 |
parameters will be made available at the node group level (in
|
|
44 |
addition to the cluster level), so that users can set defaults for
|
|
45 |
the node group::
|
|
46 |
|
|
47 |
$ gnt-group add/modify\
|
|
48 |
> -H hv:parameter=value
|
|
49 |
|
|
50 |
This changes the hypervisor inheritence level as::
|
|
51 |
|
|
52 |
cluster -> group -> OS -> instance
|
|
53 |
|
|
54 |
- Furthermore, the hypervisor parameter ``mem_path`` will be changeable
|
|
55 |
only at the cluster or node group level and users must not be able to
|
|
56 |
override this at OS or instance level. The following command must
|
|
57 |
produce an error message that ``mem_path`` may only be set at either
|
|
58 |
the cluster or the node group level::
|
|
59 |
|
|
60 |
$ gnt-instance add -H kvm:mem_path=/mount/point/for/hugepages
|
|
61 |
|
|
62 |
Memory Pools:
|
|
63 |
-------------
|
|
64 |
Memory management of Ganeti will be improved by creating separate pools
|
|
65 |
for memory used by the node itself, memory used by the hypervisor and
|
|
66 |
the memory reserved for huge pages as:
|
|
67 |
- mtotal/xen (Xen memory)
|
|
68 |
- mfree/xen (Xen unused memory)
|
|
69 |
- mtotal/hp (Memory reserved for Huge Pages)
|
|
70 |
- mfree/hp (Memory available from unused huge pages)
|
|
71 |
- mpgsize/hp (Size of a huge page)
|
|
72 |
|
|
73 |
mfree and mtotal will be changed to mean "the total and free memory for
|
|
74 |
the default method in this cluster/nodegroup". Note that the default
|
|
75 |
method depends both on the default hypervisor and its parameters.
|
|
76 |
|
|
77 |
iAllocator Changes:
|
|
78 |
-------------------
|
|
79 |
If huge pages are set as default for a cluster of node group, then
|
|
80 |
iAllocator must consider the huge pages memory on the nodes, as a
|
|
81 |
parameter when trying to find the best node for the VM.
|
|
82 |
Note that the iallocator will also be changed to use the correct
|
|
83 |
parameter depending on the cluster/group.
|
|
84 |
|
|
85 |
hbal Changes:
|
|
86 |
-------------
|
|
87 |
The cluster balancer (hbal) will be changed to use the default memory
|
|
88 |
pool and recognize memory reserved for huge pages when trying to
|
|
89 |
rebalance the cluster.
|
|
90 |
|
|
91 |
.. vim: set textwidth=72 :
|
|
92 |
.. Local Variables:
|
|
93 |
.. mode: rst
|
|
94 |
.. fill-column: 72
|
|
95 |
.. End:
|