[Gluster-users] Per-client prefered server?

Yannick Perret yannick.perret at liris.cnrs.fr
Fri Mar 4 09:13:16 UTC 2016


Le 03/03/2016 15:17, Saravanakumar Arumugam a écrit :
>
>
> On 03/03/2016 05:38 PM, Yannick Perret wrote:
>> Hello,
>>
>> I can't find if it is possible to set a prefered server on a 
>> per-client basis for replica volumes, so I ask the question here.
>>
>> The context: we have 2 storage servers, each in one building. We also 
>> have several virtual machines on each building, and they can migrate 
>> from one building to an other (depending on load, maintenance…).
>>
>> So (for testing at this time) I setup a x2 replica volume, one 
>> replica on each storage server of course. As most of our volumes are 
>> "many reads - few writes" it would be better for bandwidth that each 
>> client uses the "nearest" storage server (local building switch) - 
>> for reading, of course. The 2 buildings have a good netlink but we 
>> prefer to minimize - when not needed - data transferts beetween them 
>> (this link is shared).
>>
>> Can you see a solution for this kind of tuning? As far as I 
>> understand geo-replica is not really what I need, no?
>
> Yes, geo-replication "cannot" be used as you wish to carry out "write" 
> operation on Slave side.
>
Ok, thanks. I was pretty sure it was the case but I prefer to ask.
>>
>> It exists "cluster.read-subvolume" option of course but we can have 
>> clients on both building so a per-volume option is not what we need. 
>> An per-client equivalent of this option should be nice.
>>
>> I tested by myself a small patch to perform this - and it seems to 
>> work fine as far as I can see - but 1. before continuing in this way 
>> I would first check if it exists an other way and 2. I'm not familiar 
>> with the whole code so I'm not sure that my tests are in the 
>> "state-of-the-art" for glusterfs.
>>
> maybe you should share that interesting patch :) and get better 
> feedback about your test case.

My "patch" is quite simple: I added in 
afr_read_subvol_select_by_policy() 
(xlators/cluster/afr/src/afr-common.c) a target selection similar to the 
one managed by "read_child" configuration (see patches at the end).

Of course I also added the definition of this "forced_child" in afr.h, 
in the same way favorite_child or read_child is defined.


My real problem here is how to tell to client to change its 
"forced-child" value.

I did this by reading it from a local file 
(/var/lib/glusterd/forced-child) in init() and reconfigure() (from 
xlators/cluster/afr/src/afr.c). This is fine at startup, and when volume 
configuration changed, but I find the sending a SIGUP is not enough 
because client detects that no change occurs and do not call 
"reconfigure()". So for my tests I modified 
glusterfsd/src/glusterfsd-mgmt.c so that if 
/var/lib/glusterf/forced-child exists it behave as if a configuration 
change occured (and so calls reconfigure() which reload the forced-child 
value).


At this point it works as I expected but I think it is possible to 
handle a new forced-child value without calling all reconfigure().
Moreover this file contains a raw number, it would be better to use a 
server name, which would be converted into index after that.

If possible it may be better to send the new value using 'gluster' 
command on client? I.e. something like 'gluster volume set 
client.prefered-server SERVER-NAME'?


Any advices welcome.


Regards,
--
Y.


<<<
--- glusterfs-3.6.7/xlators/cluster/afr/src/afr-common.c 2015-11-25 
12:55:58.000000000 +0100
+++ glusterfs-3.6.7b/xlators/cluster/afr/src/afr-common.c 2015-12-10 
14:59:18.898580772 +0100
@@ -764,10 +764,18 @@
      int             i           = 0;
      int             read_subvol = -1;
      afr_private_t  *priv        = NULL;
-        afr_read_subvol_args_t local_args = {0,};
+    afr_read_subvol_args_t local_args = {0,};

      priv = this->private;

+
+    /* if forced-child use it */
+    if ((priv->forced_child >= 0)
+        && (priv->forced_child < priv->child_count)
+        && (readable[priv->forced_child])) {
+        return priv->forced_child;
+    }
+
      /* first preference - explicitly specified or local subvolume */
      if (priv->read_child >= 0 && readable[priv->read_child])
                  return priv->read_child;
 >>>

<<<
@@ -83,6 +83,7 @@
          unsigned int hash_mode;       /* for when read_child is not set */
          int favorite_child;  /* subvolume to be preferred in resolving
                                           split-brain cases */
+        int forced_child;    /* child to use (if possible) */

          gf_boolean_t inodelk_trace;
          gf_boolean_t entrylk_trace;
 >>>

<<<
--- glusterfs-3.6.7/xlators/cluster/afr/src/afr.c    2015-11-25 
12:55:58.000000000 +0100
+++ glusterfs-3.6.7b/xlators/cluster/afr/src/afr.c    2015-12-10 
16:34:55.530790442 +0100
@@ -23,6 +23,7 @@

  struct volume_options options[];

+
  int32_t
  notify (xlator_t *this, int32_t event,
          void *data, ...)
@@ -106,9 +107,26 @@
          int            ret         = -1;
          int            index       = -1;
          char          *qtype       = NULL;
+    FILE          *prefer      = NULL;
+        int            i           = -1;

          priv = this->private;

+        /* if /var/lib/glusterd/forced-child exists read the content
+           and use it as prefered target for read */
+        priv->forced_child = -1;
+        prefer = fopen("/var/lib/glusterd/forced-child", "r");
+        if (prefer) {
+                if (fscanf(prefer, "%d", &i) == 1) {
+                        if ((i >= 0) && (i < priv->child_count)) {
+                                priv->forced_child = i;
+                gf_log (this->name, GF_LOG_INFO,
+                        "using %d as forced-child", i);
+                        }
+                }
+                fclose(prefer);
+        }
+
      GF_OPTION_RECONF ("afr-dirty-xattr",
                priv->afr_dirty, options, str,
                out);
@@ -234,6 +252,7 @@
          int            read_subvol_index = -1;
          xlator_t      *fav_child   = NULL;
          char          *qtype       = NULL;
+    FILE          *prefer      = NULL;

          if (!this->children) {
                  gf_log (this->name, GF_LOG_ERROR,
@@ -261,6 +280,21 @@

          priv->read_child = -1;

+    /* if /var/lib/glusterd/forced-child exists read the content
+           and use it as prefered target for read */
+        priv->forced_child = -1;
+        prefer = fopen("/var/lib/glusterd/forced-child", "r");
+        if (prefer) {
+        if (fscanf(prefer, "%d", &i) == 1) {
+            if ((i >= 0) && (i < priv->child_count)) {
+                priv->forced_child = i;
+                gf_log (this->name, GF_LOG_INFO,
+                                        "using %d as forced-child", i);
+            }
+        }
+        fclose(prefer);
+    }
+
      GF_OPTION_INIT ("afr-dirty-xattr", priv->afr_dirty, str, out);

      GF_OPTION_INIT ("metadata-splitbrain-forced-heal",
 >>>

<<<
--- glusterfs-3.6.7/glusterfsd/src/glusterfsd-mgmt.c    2015-11-25 
12:55:58.000000000 +0100
+++ glusterfs-3.6.7b/glusterfsd/src/glusterfsd-mgmt.c    2015-12-10 
16:34:20.530789162 +0100
@@ -1502,7 +1502,9 @@
          if (size == oldvollen && (memcmp (oldvolfile, rsp.spec, size) 
== 0)) {
                  gf_log (frame->this->name, GF_LOG_INFO,
                          "No change in volfile, continuing");
-                goto out;
+        if (access("/var/lib/glusterd/forced-child", R_OK) != 0) {
+                    goto out; /* don't skip if exists to re-read 
forced-child */
+        }
          }

          tmpfp = tmpfile ();
 >>>

--
Y.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.gluster.org/pipermail/gluster-users/attachments/20160304/d666c784/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 3369 bytes
Desc: Signature cryptographique S/MIME
URL: <http://www.gluster.org/pipermail/gluster-users/attachments/20160304/d666c784/attachment.p7s>


More information about the Gluster-users mailing list