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. Thepolicies for when, where, and how much to migrate are hard-coded into themetadata balancing module. Mantle is a programmable metadata balancer builtinto the MDS. The idea is to protect the mechanisms for balancing load(migration, replication, fragmentation) but stub out the balancing policiesusing Lua. Mantle is based on [1] but the current implementation does _NOT_have the following features from that paper:

  • Balancing API: in the paper, the user fills in when, where, how much, andload calculation policies; currently, Mantle only requires that Lua policiesreturn a table of target loads (e.g., how much load to send to each MDS)

  • “How much” hook: in the paper, there was a hook that let the user controlthe fragment selector policy; currently, Mantle does not have this hook

  • 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 daemonsand clients on one node can overload the system. Let it run for a while, eventhough 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 youcannot 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 weneed to generate enough load to get over the MIN_OFFLOAD threshold that isarbitrarily set in the balancer. For example, this does not create enoughmetadata load:

  1. while true; do
  2. touch "/cephfs/blah-`date`"
  3. done

Mantle with vstart.sh

  • Start Ceph and tune the logging so we can see migrations happen:
  1. cd build
  2. ../src/vstart.sh -n -l
  3. for i in a b c; do
  4. bin/ceph --admin-daemon out/mds.$i.asok config set debug_ms 0
  5. bin/ceph --admin-daemon out/mds.$i.asok config set debug_mds 2
  6. bin/ceph --admin-daemon out/mds.$i.asok config set mds_beacon_grace 1500
  7. done
  • Put the balancer into RADOS:
  1. bin/rados put --pool=cephfs_metadata_a greedyspill.lua ../src/mds/balancers/greedyspill.lua
  • Activate Mantle:
  1. bin/ceph fs set cephfs max_mds 5
  2. bin/ceph fs set cephfs_a balancer greedyspill.lua
  • Mount CephFS in another window:
  1. bin/ceph-fuse /cephfs -o allow_other &
  2. tail -f out/mds.a.log
  3.  
  4.  
  5. Note that if you look at the last MDS (which could be a, b, or c -- it's
  6. random), you will see an an attempt to index a nil value. This is because the
  7. last MDS tries to check the load of its neighbor, which does not exist.
  • Run a simple benchmark. In our case, we use the Docker mdtest image tocreate load:
  1. for i in 0 1 2; do
  2. docker run -d \
  3. --name=client$i \
  4. -v /cephfs:/cephfs \
  5. michaelsevilla/mdtest \
  6. -F -C -n 100000 -d "/cephfs/client-test$i"
  7. done
  • When you are done, you can kill all the clients with:
  1. for i in 0 1 2 3; do docker rm -f client$i; done

Output

Looking at the log for the first MDS (could be a, b, or c), we see thateveryone has no load:

  1. 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
  2. 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
  3. 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
  4. 2016-08-21 06:44:01.764010 7fd03aaf7700 2 lua.balancer when: not migrating! my_load=0.0 hisload=0.0
  5. 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 spillbalancer dictates that half the load goes to your neighbor MDS, so we see thatMantle tries to send 1953 load units to MDS1.

  1. 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
  2. 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
  3. 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
  4. 2016-08-21 06:45:21.870034 7fd03aaf7700 2 lua.balancer when: migrating! my_load=1953.3492228857 hisload=0.0
  5. 2016-08-21 06:45:21.870050 7fd03aaf7700 2 mds.0.bal mantle decided that new targets={0=0,1=976.675,2=0}
  6. 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]
  7. 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:

  1. 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
  2. 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
  3. 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
  4. 2016-08-21 06:47:10.210298 7fd03aaf7700 2 lua.balancer when: not migrating! my_load=415.79000078186 hisload=186.5606496623
  5. 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 balancerpolicies 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 aCeph CLI command (“ceph fs set cephfs balancer mybalancer.lua”). If the Lua policyfails (for whatever reason), we fall back to the original metadata loadbalancer. The balancer is stored in the RADOS metadata pool and a string in theMDSMap tells the MDSs which balancer to use.

Exposing Metrics to Lua

Metrics are exposed directly to the Lua code as global variables instead ofusing a well-defined function signature. There is a global “mds” table, whereeach index is an MDS number (e.g., 0) and each value is a dictionary of metricsand values. The Lua code can grab metrics using something like this:

  1. 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 easierto add new metrics without having to change the API on the Lua side; we wantthe API to grow and shrink as we explore which metrics matter. The downside ofthis approach is that the person programming Lua balancer policies has to lookat the Ceph source code to see which metrics are exposed. We figure that theMantle developer will be in touch with MDS internals anyways.

The metrics exposed to the Lua policy are the same ones that are already storedin 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 errorsin the MDBalancer. We do not want the error propagating up the call chain. Thecls_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 beleveraged 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 tothe amount of load to send to each MDS. These loads are inserted directly intothe MDBalancer “my_targets” vector. We do not allow the MDS to return a tableof MDSs and metrics because we want the decision to be completely made on theLua side.

Iterating through tables returned by Lua is done through the stack. In Luajargon: a dummy value is pushed onto the stack and the next iterator replacesthe top of the stack with a (k, v) pair. After reading each value, pop thatvalue 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 changesin the MDS Map. The balancer pulls the Lua code from RADOS synchronously. We dothis with a timeout: if the asynchronous read does not come back within halfthe balancing tick interval the operation is cancelled and a Connection Timeouterror is returned. By default, the balancing tick interval is 10 seconds, soMantle will use a 5 second second timeout. This design allows Mantle toimmediately return an error if anything RADOS-related goes wrong.

We use this implementation because we do not want to do a blocking OSD readfrom inside the global MDS lock. Doing so would bring down the MDS cluster ifany of the OSDs are not responsive – this is tested in the ceph-qa-suite bysetting 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 Mapand fill in the Lua code in the background. We cannot do this because the MDSdoes not support daemon-local fallbacks and the balancer assumes that all MDSscome 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 asthe cls logging interface:

  1. 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 isactually calling the dout function in C++.

Warning and Info messages are centralized using the clog/Beacon. Successfulmessages are only sent on version changes by the first MDS to avoid spammingthe 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 nottest invalid balancer logging and loading the actual Lua VM.