[sheepdog-users] What is cooking in master # 2014.1.22

[Liu Yuan]

Hello sheepdog walkers,

This is the report of current status of development of sheepdog project.

This is a summery of what is new in v0.8.0 compared to v0.7.0

User-visible changes:

 1 Data distribution algorithm is changed
   We introduce a better hash algorithm for data distribution over all nodes and
 disks. As a result, you'll expect better data and request distribution over
 nodes and disks. This change is NOT backward compatibable, meaning that user
 can't upgrade from v0.7.x to v0.8.0 directly. One method to migrate old data
 to v0.8.x is use 'dog cluster snapshot' to save cluster data first by v0.7.x
 and then load into v0.8.0.

 2 add x86 hardware acceleration for sha1
   Cluster snapshot and recovery rely on the sha1 algorithm, so with this
 acceleration, users will expect a slightly faster recovery/snapshoting if your
 CPU support ssse3.

 3 Basic request statistics
   A new 'dog node stat' command added to give some sort of statistics.

   $ dog node stat -w

    Request     Active  Total   Write   Read    Remove  Flush   All WR  All RD  WRBW    RDBW    RPS
    Client      5       341362  100459  234375  0       6528    44 GB   26 GB   98 MB   0.0 MB  0
    Peer        0       41128   4008    37120   0       0       8.6 GB  3.9 GB  24 MB   0.0 MB  0
    Request     Active  Total   Write   Read    Remove  Flush   All WR  All RD  WRBW    RDBW    RPS
    Client      0       341969  100459  234982  0       6528    44 GB   26 GB   92 MB   0.0 MB  0
    Peer        0       41220   4008    37212   0       0       8.6 GB  3.9 GB  24 MB   0.0 MB  0

   Client means all the requests from VMs on this node.
   Peer means all the requests from other nodes
   Active: active request being handled for now
   RPS: Reuqest per second
   BW: bandwidth per second

 4 Multi-Disk enhancement
   Now we support unlimited number of disks in local nodes. Previously, we
   support at most 64 disks.

 5 Maximum number of cluster
   With zookeeper, we can support 6k+ nodes right now. Previously, we suppport
   1024 nodes at most.

 6 Erasure Code
   Erasure Coding(EC) is a redundancy scheme that achieves high available of
   data with much less storage overhead compared to complete replication. EC
   in sheepdog is supposed to run VM guest, not only for cold data, because our
   EC's performance is somewhat better than full replication. More info, see

   https://github.com/sheepdog/sheepdog/wiki/Erasure-Code-Support

 7 Hyper Volume
   We introduce hyper volume (up to 16 PB) to this release. Currently, only
   sheepfs and http service make use of it, this means you can mount a 16PB
   volume into local file system via sheepfs. QEMU and iSCSI doesn't support
   it yet.

 8 HTTP Simple Storage
   Sheepdog HTTP simple storage provides simple object store/retrieve service 
   via RESTful API, similar to Openstack Swift or Amazone S3. Our HTTP API only
   support GET/PUT/DELETE/POST/HEAD operation and we implement a subset of
   Openstack Swift API for now and plan to be Amazone S3 compatible in the
   future. 
   
   We have a same API to upload/download small/big objects and support object
   size upto 16PB.

   More info, see
  
   https://github.com/sheepdog/sheepdog/wiki/HTTP-Simple-Storage

 9 Logger Enhancement
   Now user can rotate sheep log(sheep.log) by signal SIGHUP and 
   we also support log dog operaton by setting SHEEPDOG_DOG_LOG and
   SHEEPDOG_DOG_LOG_PATH for dog. E.g,

   $ export SHEEPDOG_DOG_LOG
   $ export SHEEPDOG_DOG_LOG_PATH=/var/log/dog.log # if not set, we'll log in syslog
   # then all your dog operations are logged.

 10 strict mode added back
   Previously, we support safe/qurum/unsafe mode, but removed in the later
   commit because for some corner case we can't keep 'safe' promise.

   But sometimes we want to make sure we write the exact number of copies to
   honor the promise of the redundancy for "strict mode". This means that after
   writing of targeted data, they are redundant as promised and can withstand
   the random node failures.

   For example, with a 4:2 policy, we need at least write to 6 nodes with data
   strip and parity strips. For non-strict mode, we allow to write successfully
   only if the data are written fully with 4 nodes alive.

   We can pass '-t|--strict' for 'dog cluster format' to enable strict mode.

 11 Sheepdog VM auto-reconnect is support officially by QEMU v1.7
   Because this is all about the codes in QEMU sheepdog block driver,
   auto-reconnect feature is supported by all the sheepdog release.

 12 Synchronous 'dog vdi delete'
   This command will wait for all the data are removed before return. Previously
   we return immediately even if 'dog vdi delete' fails after some time later.


*******************************************************************************
Old posts of the report:

  What is cooking in master # 2013.5.23
  http://lists.wpkg.org/pipermail/sheepdog-users/2013-May/000746.html
   
  What is cooking in master # 2013.8.9
  http://lists.wpkg.org/pipermail/sheepdog-users/2013-August/001142.html

Thanks
Yuan