diff options
author | Jeff Darcy <jdarcy@redhat.com> | 2012-03-12 09:32:40 -0400 |
---|---|---|
committer | Anand Avati <avati@redhat.com> | 2012-05-31 17:29:01 -0700 |
commit | ddc044bfa2840981de4003c3b9efcac84387dc2b (patch) | |
tree | a83d476702cac7ecc7ae59057c368f622a51af4c /xlators/cluster/afr/src/afr.h | |
parent | e066a5fea7bdaa5da78e49c9a5bf344af2f33d3c (diff) |
replicate: add hashed read-child method.
Both the first-to-respond method and the round-robin method are susceptible
to clients repeatedly choosing the same servers across a series of opens,
creating hot spots. Also, the code to handle a replica being down will
ignore both methods and just choose the first remaining (which is not an
issue for two-way but can be otherwise). The hashed method more reliably
avoids such hot spots. There are three values/modes.
0: use the old (broken) methods.
1: select a read-child based on a hash of the file's GFID, so all clients
will choose the same subvolume for a file (ensuring maximum consistency)
but will distribute load for a set of files.
2: select a read-child based on a hash of the file's GFID plus the client's
PID, so different children will distribute load even for one file.
Mode 2 will probably be optimal for most cases. Using response time when we
open the file is problematic, both because a single sample might not have
been representative even then and because load might have shifted in the
hours or days since (for long-lived files). Trying to use more current load
information can lead to "herd following" behavior which is just as bad.
Pseudo-random distribution is likely to be the best we can reasonably do,
just as it is for DHT.
Change-Id: I798c2760411eacf32e82a85f03bb7b08a4a49461
BUG: 802513
Signed-off-by: Jeff Darcy <jdarcy@redhat.com>
Reviewed-on: http://review.gluster.com/2926
Tested-by: Gluster Build System <jenkins@build.gluster.com>
Reviewed-by: Anand Avati <avati@redhat.com>
Diffstat (limited to 'xlators/cluster/afr/src/afr.h')
-rw-r--r-- | xlators/cluster/afr/src/afr.h | 6 |
1 files changed, 4 insertions, 2 deletions
diff --git a/xlators/cluster/afr/src/afr.h b/xlators/cluster/afr/src/afr.h index 27cb83a5791..a1a30562bf1 100644 --- a/xlators/cluster/afr/src/afr.h +++ b/xlators/cluster/afr/src/afr.h @@ -130,6 +130,7 @@ typedef struct _afr_private { gf_boolean_t entry_change_log; /* on/off */ int read_child; /* read-subvolume */ + unsigned int hash_mode; /* for when read_child is not set */ int favorite_child; /* subvolume to be preferred in resolving split-brain cases */ @@ -936,12 +937,13 @@ afr_first_up_child (unsigned char *child_up, size_t child_count); int afr_select_read_child_from_policy (int32_t *fresh_children, int32_t child_count, int32_t prev_read_child, - int32_t config_read_child, int32_t *sources); + int32_t config_read_child, int32_t *sources, + unsigned int hmode, uuid_t gfid); void afr_set_read_ctx_from_policy (xlator_t *this, inode_t *inode, int32_t *fresh_children, int32_t prev_read_child, - int32_t config_read_child); + int32_t config_read_child, uuid_t gfid); int32_t afr_get_call_child (xlator_t *this, unsigned char *child_up, int32_t read_child, |