Mantle
======
.. warning::
Mantle is for research and development of metadata balancer algorithms,
not for use on production CephFS clusters.
Multiple, active MDSs can migrate directories to balance metadata load. The
policies for when, where, and how much to migrate are hard-coded into the
metadata balancing module. Mantle is a programmable metadata balancer built
into the MDS. The idea is to protect the mechanisms for balancing load
(migration, replication, fragmentation) but stub out the balancing policies
using Lua. Mantle is based on [1] but the current implementation does *NOT*
have the following features from that paper:
1. Balancing API: in the paper, the user fills in when, where, how much, and
load calculation policies; currently, Mantle only requires that Lua policies
return a table of target loads (e.g., how much load to send to each MDS)
2. "How much" hook: in the paper, there was a hook that let the user control
the fragment selector policy; currently, Mantle does not have this hook
3. Instantaneous CPU utilization as a metric
[1] Supercomputing '15 Paper:
http://sc15.supercomputing.org/schedule/event_detail-evid=pap168.html
Quickstart with vstart
----------------------
.. warning::
Developing balancers with vstart is difficult because running all daemons
and clients on one node can overload the system. Let it run for a while, even
though you will likely see a bunch of lost heartbeat and laggy MDS warnings.
Most of the time this guide will work but sometimes all MDSs lock up and you
cannot actually see them spill. It is much better to run this on a cluster.
As a prerequisite, we assume you have installed `mdtest
`_ or pulled the `Docker image
`_. We use mdtest because we
need to generate enough load to get over the MIN_OFFLOAD threshold that is
arbitrarily set in the balancer. For example, this does not create enough
metadata load:
::
while true; do
touch "/cephfs/blah-`date`"
done
Mantle with `vstart.sh`
~~~~~~~~~~~~~~~~~~~~~~~
1. Start Ceph and tune the logging so we can see migrations happen:
::
cd build
../src/vstart.sh -n -l
for i in a b c; do
bin/ceph --admin-daemon out/mds.$i.asok config set debug_ms 0
bin/ceph --admin-daemon out/mds.$i.asok config set debug_mds 2
bin/ceph --admin-daemon out/mds.$i.asok config set mds_beacon_grace 1500
done
2. Put the balancer into RADOS:
::
bin/rados put --pool=cephfs_metadata_a greedyspill.lua ../src/mds/balancers/greedyspill.lua
3. Activate Mantle:
::
bin/ceph fs set cephfs max_mds 5
bin/ceph fs set cephfs_a balancer greedyspill.lua
4. Mount CephFS in another window:
::
bin/ceph-fuse /cephfs -o allow_other &
tail -f out/mds.a.log
Note that if you look at the last MDS (which could be a, b, or c -- it's
random), you will see an an attempt to index a nil value. This is because the
last MDS tries to check the load of its neighbor, which does not exist.
5. Run a simple benchmark. In our case, we use the Docker mdtest image to
create load:
::
for i in 0 1 2; do
docker run -d \
--name=client$i \
-v /cephfs:/cephfs \
michaelsevilla/mdtest \
-F -C -n 100000 -d "/cephfs/client-test$i"
done
6. When you are done, you can kill all the clients with:
::
for i in 0 1 2 3; do docker rm -f client$i; done
Output
~~~~~~
Looking at the log for the first MDS (could be a, b, or c), we see that
everyone has no load:
::
2016-08-21 06:44:01.763930 7fd03aaf7700 0 lua.balancer MDS0: < auth.meta_load=0.0 all.meta_load=0.0 req_rate=1.0 queue_len=0.0 cpu_load_avg=1.35 > load=0.0
2016-08-21 06:44:01.763966 7fd03aaf7700 0 lua.balancer MDS1: < auth.meta_load=0.0 all.meta_load=0.0 req_rate=0.0 queue_len=0.0 cpu_load_avg=1.35 > load=0.0
2016-08-21 06:44:01.763982 7fd03aaf7700 0 lua.balancer MDS2: < auth.meta_load=0.0 all.meta_load=0.0 req_rate=0.0 queue_len=0.0 cpu_load_avg=1.35 > load=0.0
2016-08-21 06:44:01.764010 7fd03aaf7700 2 lua.balancer when: not migrating! my_load=0.0 hisload=0.0
2016-08-21 06:44:01.764033 7fd03aaf7700 2 mds.0.bal mantle decided that new targets={}
After the jobs starts, MDS0 gets about 1953 units of load. The greedy spill
balancer dictates that half the load goes to your neighbor MDS, so we see that
Mantle tries to send 1953 load units to MDS1.
::
2016-08-21 06:45:21.869994 7fd03aaf7700 0 lua.balancer MDS0: < auth.meta_load=5834.188908912 all.meta_load=1953.3492228857 req_rate=12591.0 queue_len=1075.0 cpu_load_avg=3.05 > load=1953.3492228857
2016-08-21 06:45:21.870017 7fd03aaf7700 0 lua.balancer MDS1: < auth.meta_load=0.0 all.meta_load=0.0 req_rate=0.0 queue_len=0.0 cpu_load_avg=3.05 > load=0.0
2016-08-21 06:45:21.870027 7fd03aaf7700 0 lua.balancer MDS2: < auth.meta_load=0.0 all.meta_load=0.0 req_rate=0.0 queue_len=0.0 cpu_load_avg=3.05 > load=0.0
2016-08-21 06:45:21.870034 7fd03aaf7700 2 lua.balancer when: migrating! my_load=1953.3492228857 hisload=0.0
2016-08-21 06:45:21.870050 7fd03aaf7700 2 mds.0.bal mantle decided that new targets={0=0,1=976.675,2=0}
2016-08-21 06:45:21.870094 7fd03aaf7700 0 mds.0.bal - exporting [0,0.52287 1.04574] 1030.88 to mds.1 [dir 100000006ab /client-test2/ [2,head] auth pv=33 v=32 cv=32/0 ap=2+3+4 state=1610612802|complete f(v0 m2016-08-21 06:44:20.366935 1=0+1) n(v2 rc2016-08-21 06:44:30.946816 3790=3788+2) hs=1+0,ss=0+0 dirty=1 | child=1 dirty=1 authpin=1 0x55d2762fd690]
2016-08-21 06:45:21.870151 7fd03aaf7700 0 mds.0.migrator nicely exporting to mds.1 [dir 100000006ab /client-test2/ [2,head] auth pv=33 v=32 cv=32/0 ap=2+3+4 state=1610612802|complete f(v0 m2016-08-21 06:44:20.366935 1=0+1) n(v2 rc2016-08-21 06:44:30.946816 3790=3788+2) hs=1+0,ss=0+0 dirty=1 | child=1 dirty=1 authpin=1 0x55d2762fd690]
Eventually load moves around:
::
2016-08-21 06:47:10.210253 7fd03aaf7700 0 lua.balancer MDS0: < auth.meta_load=415.77414300449 all.meta_load=415.79000078186 req_rate=82813.0 queue_len=0.0 cpu_load_avg=11.97 > load=415.79000078186
2016-08-21 06:47:10.210277 7fd03aaf7700 0 lua.balancer MDS1: < auth.meta_load=228.72023977691 all.meta_load=186.5606496623 req_rate=28580.0 queue_len=0.0 cpu_load_avg=11.97 > load=186.5606496623
2016-08-21 06:47:10.210290 7fd03aaf7700 0 lua.balancer MDS2: < auth.meta_load=0.0 all.meta_load=0.0 req_rate=1.0 queue_len=0.0 cpu_load_avg=11.97 > load=0.0
2016-08-21 06:47:10.210298 7fd03aaf7700 2 lua.balancer when: not migrating! my_load=415.79000078186 hisload=186.5606496623
2016-08-21 06:47:10.210311 7fd03aaf7700 2 mds.0.bal mantle decided that new targets={}
Implementation Details
----------------------
Most of the implementation is in MDBalancer. Metrics are passed to the balancer
policies via the Lua stack and a list of loads is returned back to MDBalancer.
It sits alongside the current balancer implementation and it's enabled with a
Ceph CLI command ("ceph fs set cephfs balancer mybalancer.lua"). If the Lua policy
fails (for whatever reason), we fall back to the original metadata load
balancer. The balancer is stored in the RADOS metadata pool and a string in the
MDSMap tells the MDSs which balancer to use.
Exposing Metrics to Lua
~~~~~~~~~~~~~~~~~~~~~~~
Metrics are exposed directly to the Lua code as global variables instead of
using a well-defined function signature. There is a global "mds" table, where
each index is an MDS number (e.g., 0) and each value is a dictionary of metrics
and values. The Lua code can grab metrics using something like this:
::
mds[0]["queue_len"]
This is in contrast to cls-lua in the OSDs, which has well-defined arguments
(e.g., input/output bufferlists). Exposing the metrics directly makes it easier
to add new metrics without having to change the API on the Lua side; we want
the API to grow and shrink as we explore which metrics matter. The downside of
this approach is that the person programming Lua balancer policies has to look
at the Ceph source code to see which metrics are exposed. We figure that the
Mantle developer will be in touch with MDS internals anyways.
The metrics exposed to the Lua policy are the same ones that are already stored
in mds_load_t: auth.meta_load(), all.meta_load(), req_rate, queue_length,
cpu_load_avg.
Compile/Execute the Balancer
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Here we use `lua_pcall` instead of `lua_call` because we want to handle errors
in the MDBalancer. We do not want the error propagating up the call chain. The
cls_lua class wants to handle the error itself because it must fail gracefully.
For Mantle, we don't care if a Lua error crashes our balancer -- in that case,
we will fall back to the original balancer.
The performance improvement of using `lua_call` over `lua_pcall` would not be
leveraged here because the balancer is invoked every 10 seconds by default.
Returning Policy Decision to C++
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We force the Lua policy engine to return a table of values, corresponding to
the amount of load to send to each MDS. These loads are inserted directly into
the MDBalancer "my_targets" vector. We do not allow the MDS to return a table
of MDSs and metrics because we want the decision to be completely made on the
Lua side.
Iterating through tables returned by Lua is done through the stack. In Lua
jargon: a dummy value is pushed onto the stack and the next iterator replaces
the top of the stack with a (k, v) pair. After reading each value, pop that
value but keep the key for the next call to `lua_next`.
Reading from RADOS
~~~~~~~~~~~~~~~~~~
All MDSs will read balancing code from RADOS when the balancer version changes
in the MDS Map. The balancer pulls the Lua code from RADOS synchronously. We do
this with a timeout: if the asynchronous read does not come back within half
the balancing tick interval the operation is cancelled and a Connection Timeout
error is returned. By default, the balancing tick interval is 10 seconds, so
Mantle will use a 5 second second timeout. This design allows Mantle to
immediately return an error if anything RADOS-related goes wrong.
We use this implementation because we do not want to do a blocking OSD read
from inside the global MDS lock. Doing so would bring down the MDS cluster if
any of the OSDs are not responsive -- this is tested in the ceph-qa-suite by
setting all OSDs to down/out and making sure the MDS cluster stays active.
One approach would be to asynchronously fire the read when handling the MDS Map
and fill in the Lua code in the background. We cannot do this because the MDS
does not support daemon-local fallbacks and the balancer assumes that all MDSs
come to the same decision at the same time (e.g., importers, exporters, etc.).
Debugging
~~~~~~~~~
Logging in a Lua policy will appear in the MDS log. The syntax is the same as
the cls logging interface:
::
BAL_LOG(0, "this is a log message")
It is implemented by passing a function that wraps the `dout` logging framework
(`dout_wrapper`) to Lua with the `lua_register()` primitive. The Lua code is
actually calling the `dout` function in C++.
Warning and Info messages are centralized using the clog/Beacon. Successful
messages are only sent on version changes by the first MDS to avoid spamming
the `ceph -w` utility. These messages are used for the integration tests.
Testing
~~~~~~~
Testing is done with the ceph-qa-suite (tasks.cephfs.test_mantle). We do not
test invalid balancer logging and loading the actual Lua VM.