glusterfs.git/xlators/cluster/dht/src/tier.h, branch round-robin2

cluster/tier: make promotion and demotion independant

2016-02-29T01:32:22+00:00

Currently a main loop in tiering spawns promotion and demotion threads,
and does a join to wait for them to complete. When one of the two
threads takes a long time, the main thread waits for it
before exiting the join. It may wait so long the scheduled time
for the other thread is skipped. In the case of demotion, it may be
a long time before another attempt. This patch fixes that by
making the promotion and demotion activities independant. A side
effect of this change is the logic is significantly simplified.

Change-Id: I1196bd4bbfc95e8aa326a9bd4ebf395032369d1c
BUG: 1306852
Signed-off-by: Dan Lambright 
Reviewed-on: http://review.gluster.org/13433
Smoke: Gluster Build System 
Reviewed-by: Joseph Fernandes
NetBSD-regression: NetBSD Build System 
CentOS-regression: Gluster Build System

tier/gfdb : Round-Robin read of query files

2016-02-03T18:04:40+00:00

1. Each brick on a host will get a separate query file.
2. While reading query record from these query files we
   read them in a Round-Robin manner.
3. When an error occurs during migration we rename it to
   query file with an time stamp and .err extension for
   better debugging.

Change-Id: I27c4285d24fd695d2d5cbd9fd7db3879d277ecc8
BUG: 1302772
Signed-off-by: Joseph Fernandes 
Reviewed-on: http://review.gluster.org/13293
Smoke: Gluster Build System 
Tested-by: N Balachandran 
NetBSD-regression: NetBSD Build System 
Reviewed-by: Dan Lambright 
CentOS-regression: Gluster Build System

cluster/tier: readdirp to cold tier only

2015-11-23T12:05:55+00:00

It is possible a file would get migrated in the middle
of a readdir operation. If there are four subvolumes A,B,C,D,
and if readdir reads them in order and reaches subvol B,
then, if a file is moved from D to A, it will not be included
in the readdir output.

This phenonema has pre-existed in DHT migration but is more
apparent in tiering.

When a file is moved off the hashed subvolume a T file is created.
For tiering, we will make the cold subvolume the hashed subvolume.
This will ensure the creation of a T file. Readdir will not skip T
files in the tier translator.

Making the cold subvolume the hashed subvolume ensures the T
files created on promotions or creates will be less likely to
fill the volume.

Creates still put the data on the hot subvolume.

Change-Id: Ifde557d3d0e94a4570ca9f115adee3db2ee75407
BUG:  1281598
Signed-off-by: Dan Lambright 
Reviewed-on: http://review.gluster.org/12530
Tested-by: Gluster Build System 
Tested-by: NetBSD Build System 
Reviewed-by: N Balachandran 
Reviewed-by: Raghavendra G

cluster/tier: Do not delete linkto file on demotion

2015-11-18T15:16:12+00:00

The current DHT migration code will always delete the
src linkto file after migration as dht always moves
 files to the hashed subvol. This is not the case in tiering.
The lack of linkto files causes rename to fail leaving 2 files
 with the same name but different gfids on the volume.
Modified to leave the linkto file behind if the source
 volume is the hashed subvolume.

Change-Id: I2b99f7d34b4b719aee6232dc40c6a8f8ba88225d
BUG: 1279376
Signed-off-by: N Balachandran 
Reviewed-on: http://review.gluster.org/12551
Reviewed-by: Dan Lambright 
Tested-by: Dan Lambright

tier/libgfdb: Replacing ASCII query file with binary

2015-11-06T18:30:45+00:00

Earlier, when the database was queried we used to save
all the queried records in an ASCII format in the query file.
This caused issues like filename having ASCII delimiter and used
to take a lot of space. The tier.c file also had a lot of parsing code.

Here we changed the format of the query file to binary.
All the logic of serialization and formating of query record is done
by libgfdb. Libgfdb provides API,
gfdb_write_query_record() and gfdb_read_query_record(),
which the user i.e tier migrator and CTR xlator can use to
write to and read from query file.
With this binary format we save on disk space i.e reduce to 50% atleast
as we are saving GFID's in binary format 16 bytes and not the string format
which takes 36 bytes + We are not saving path of the file + we are also saving on
ASCII delimiters.

The on disk format of query record is as follows,

+---------------------------------------------------------------------------+
| Length of serialized query record |       Serialized Query Record         |
+---------------------------------------------------------------------------+
             4 bytes                     Length of serialized query record
                                                      |
                                                      |
     -------------------------------------------------|
     |
     |
     V
   Serialized Query Record Format:
   +---------------------------------------------------------------------------+
   | GFID |  Link count   |    |.....                      | FOOTER |
   +---------------------------------------------------------------------------+
     16 B        4 B         Link Length                                  4 B
                                |                                          |
                                |                                          |
   -----------------------------|                                          |
   |                                                                       |
   |                                                                       |
   V                                                                       |
   Each  will be serialized as                                  |
   +-----------------------------------------------+                       |
   | PGID | BASE_NAME_LENGTH |      BASE_NAME      |                       |
   +-----------------------------------------------+                       |
     16 B       4 B             BASE_NAME_LENGTH                           |
                                                                           |
                                                                           |
   ------------------------------------------------------------------------|
   |
   |
   V
   FOOTER is a magic number 0xBAADF00D indicating the end of the record.
   This also serves as a serialized schema validator.

Change-Id: I9db7416fd421e118dd44eafab8b535caafe50d5a
BUG: 1272207
Signed-off-by: Joseph Fernandes 
Reviewed-on: http://review.gluster.org/12354
Reviewed-by: N Balachandran 
Tested-by: Gluster Build System 
Reviewed-by: Dan Lambright 
Tested-by: Dan Lambright

cluster/tier: add watermarks and policy driver

2015-10-10T13:00:04+00:00

This fix introduces infrastructure to support different
policies for promotion and demotion.

Currently the tier feature automatically promotes and demotes
files periodically based on access. This is good for testing
but too stringent for most real workloads. It makes it
difficult to fully utilize a hot tier- data will be demoted
before it is touched- its unlikely a 100GB hot SSD will have
all its data touched in a window of time.

A new parameter "mode" allows the user to pick promotion/demotion
polcies.

The "test mode" will be used for *.t and other general testing.
This is the current mechanism.

The "cache mode" introduces watermarks. The watermarks
represent levels of data residing on the hot tier.

"cache mode" policy:

The % the hot tier is full is called P.

Do not promote or demote more than D MB or F files.

A random number [0-100] is called R.

Rules for migration:

if (P < watermark_low) don't demote, always promote.

if (P >= watermark_low) && (P < watermark_hi) demote if R < P; promote if R > P.

if (P > watermark_hi) always demote, don't promote.

gluster volume set {vol} cluster.watermark-hi %
gluster volume set {vol} cluster.watermark-low %
gluster volume set {vol} cluster.tier-max-mb {D}
gluster volume set {vol} cluster.tier-max-files {F}
gluster volume set {vol} cluster.tier-mode {test|cache}

Change-Id: I157f19667ec95aa1d53406041c1e3b073be127c2
BUG: 1257911
Signed-off-by: Dan Lambright 
Reviewed-on: http://review.gluster.org/12039
Tested-by: Gluster Build System 
Reviewed-by: Atin Mukherjee

cluster/tier fix bug with sql includes introduced by 12031

2015-09-12T00:12:55+00:00

We accidentally introduced a bug where client translators have a
dependency on sql. This broke freebsd smoke tests. Fix is to
abstract from the client those dependencies.

Change-Id: I7152573a489bacc8f32e6eb139f9ff4408288f5b
BUG: 1260730
Signed-off-by: Dan Lambright 
Reviewed-on: http://review.gluster.org/12155
Tested-by: NetBSD Build System

tier/ctr: Solving DB Lock issue due to write contention from db connections

2015-09-08T12:13:00+00:00

Problem: The DB on the brick is been accessed by CTR, for write and
tier migrator, for read and write. The write from tier migrator is reseting
the heat counters after a cycle. Since we are using sqlite, two connections
trying to write would cause a db lock contention. As a result CTR used to fail
to update the db.

Solution: Using the same db connection of CTR for reseting the heat counters.
1) Introducted a new IPC FOP for CTR
2) After the query do a ipc syncop to the underlying client xlator associated
   to the brick.
3) CTR in brick will catch the IPC FOP and cleat the heat counters.

Change-Id: I53306bfc08dcdba479deb4ccc154896521336150
BUG: 1260730
Signed-off-by: Joseph Fernandes 
Reviewed-on: http://review.gluster.org/12031
Tested-by: NetBSD Build System 
Tested-by: Gluster Build System 
Reviewed-by: Dan Lambright 
Tested-by: Dan Lambright

tier/ctr: Ignore creation of T file and Ctr Lookup heal improvememnts

2015-06-26T23:05:20+00:00

1) Ignore creation of T file in ctr_mknod
2) Ignore lookup for T file in ctr_lookup
3) Ctr_lookup:
    a. If the gfid and pgfid in empty dont record
    b. Decreased log level for multiple heal attempts
    c. Inode/File heal happens after an expiry period, which is configurable.
    d. Hardlink heal happens after an expiry period, which is configurable.

Change-Id: Id8eb5092e78beaec22d05f5283645081619e2452
BUG: 1235269
Signed-off-by: Joseph Fernandes 
Reviewed-on: http://review.gluster.org/11334
Tested-by: Gluster Build System 
Reviewed-by: Dan Lambright 
Tested-by: Dan Lambright

build: do not #include "config.h" in each file

2015-05-29T12:43:17+00:00

Instead of including config.h in each file, and have the additional
config.h included from the compiler commandline (-include option).

When a .c file tests for a certain #define, and config.h was not
included, incorrect assumtions were made. With this change, it can not
happen again.

BUG: 1222319
Change-Id: I4f9097b8740b81ecfe8b218d52ca50361f74cb64
Signed-off-by: Niels de Vos 
Reviewed-on: http://review.gluster.org/10808
Tested-by: Gluster Build System 
Tested-by: NetBSD Build System
Reviewed-by: Kaleb KEITHLEY 
Reviewed-by: Pranith Kumar Karampuri