Bug 1259572

+++ This bug was initially created as a clone of Bug #1255110 +++

Description of problem:
Using a replica 2 + arbiter 1 configuration for ovirt storage domain. When arbiter node is up client IO is decreased by ~%30. Monitoring bandwidth on arbiter node shows significant rx network, in this case 50-60MB/s but the brick local path on the arbiter node is not showing significant disk space usage (< 1MB).

Version-Release number of selected component (if applicable):
CentOS 6.7 / 7.1
Gluster 3.7.3

How reproducible:
Always. If Arbiter node is killed, client IO is higher.
Initially discovered using ovirt, but also easily reproduced by writing to client fuse mount point.

Steps to Reproduce:
1. Create gluster volume replica 2 arbiter 1
2. Write data on client fuse mount point
3. Watch realtime network bandwidth on arbiter node

Actual results:
Client is sending IO writes to arbiter node, decreasing expected performance.

Expected results:
No heavy IO should be going to the arbiter node, as it has no reason to receive data when it doesn't have any storage bricks. This considerably slows client IO as it is writing to 3 nodes instead of two. I would assume this is the same performance penalty as replica 3 would be vs replica 2.

Additional info:
During a disk migration from an NFS storage domain to a gluster storage domain, the arbiter node interface shows 37GB of data received:

eth1: flags=4163<UP,BROADCAST,RUNNING,MULTICAST>  mtu 1500
        inet 10.0.231.62  netmask 255.255.255.0  broadcast 10.0.231.255
        inet6 fe80::5054:ff:fe61:a934  prefixlen 64  scopeid 0x20<link>
        ether 52:54:00:61:a9:34  txqueuelen 1000  (Ethernet)
        RX packets 5874053  bytes 39820122925 (37.0 GiB)
        RX errors 0  dropped 650  overruns 0  frame 0
        TX packets 4793230  bytes 4387154708 (4.0 GiB)
        TX errors 0  dropped 0 overruns 0  carrier 0  collisions 0

BUT the arbiter node has a very little storage space available to it ('brick' mount point is on /):

# df -h
Filesystem      Size  Used Avail Use% Mounted on
/dev/vda3       8.6G  1.2G  7.4G  14% /
devtmpfs        912M     0  912M   0% /dev
tmpfs           921M     0  921M   0% /dev/shm
tmpfs           921M  8.4M  912M   1% /run
tmpfs           921M     0  921M   0% /sys/fs/cgroup
/dev/vda1       497M  157M  341M  32% /boot

--- Additional comment from Steve D on 2015-08-26 16:42:56 EDT ---

Furthermore,

If I do some write tests to a fuse mount point for the volume I get:

Active nodes: node1, node2, arbiter:
# dd if=/dev/zero of=test10 bs=1M count=200 oflag=direct
^C34+0 records in
34+0 records out
35651584 bytes (36 MB) copied, 33.394 s, 1.1 MB/s

Active nodes: node1, node2:
# dd if=/dev/zero of=test10 bs=1M count=200 oflag=direct
200+0 records in
200+0 records out
209715200 bytes (210 MB) copied, 3.19871 s, 65.6 MB/s

Active nodes: node2, arbiter
# dd if=/dev/zero of=test10 bs=1M count=200 oflag=direct
200+0 records in
200+0 records out
209715200 bytes (210 MB) copied, 5.86619 s, 35.7 MB/s
# dd if=/dev/zero of=test10 bs=1M count=200 oflag=direct
200+0 records in
200+0 records out
209715200 bytes (210 MB) copied, 7.74412 s, 27.1 MB/s
# dd if=/dev/zero of=test10 bs=1M count=200 oflag=direct
200+0 records in
200+0 records out
209715200 bytes (210 MB) copied, 6.77487 s, 31.0 MB/s

--- Additional comment from Ravishankar N on 2015-08-27 05:21:44 EDT ---

Hi Steve,
1. You're correct on the observation that we send the writes even to the brick process of the arbiter node (even though the writes are not written to the disk). But we need the write to be sent to the brick for things to work. (AFR changelog xattrs depend on this). What we could do is send only one byte to the arbiter instead of the entire data. I'll work on the fix.

2. FWIW, I'm not able to recreate the drastic difference in 'dd' throughputs (1.1 vs 65,6 MB/s ?) as described in comment #1. I notice only marginal difference in my test setup. Can you check if you're hitting the same behaviour on a normal replica 3 volume?

--- Additional comment from Steve D on 2015-09-02 15:53:37 EDT ---

Sorry took a while to get back to this. My arbiter node was a VM before, I decided to get another physical host running and added some storage.

I should also mention all my bricks are SSD's, net is 10gig.

I added a 3rd disk, replica 3 no arbiter:

Fuse mount point:
dd if=/dev/zero of=test10 bs=1M count=200 oflag=direct
200+0 records in
200+0 records out
209715200 bytes (210 MB) copied, 0.360235 s, 582 MB/s

I then rebuilt the volume with replica 3 arbiter 1:

Fuse mount point:
dd if=/dev/zero of=test bs=1M count=200 oflag=direct
200+0 records in
200+0 records out
209715200 bytes (210 MB) copied, 2.75525 s, 76.1 MB/s

Obviously there was some performance impact from the virt env, although I have no idea what that would be. But in either case there is a significant gap here.

Also I'm noticing that an Ovirt VM backed by rep3/arbiter 1 gets significantly less write speed:
dd if=/dev/zero of=test bs=1M count=200 oflag=direct
200+0 records in
200+0 records out
209715200 bytes (210 MB) copied, 16.5165 s, 12.7 MB/s

In replica 3 no arbiter, the VM gets:
dd if=/dev/zero of=test bs=1M count=200 oflag=direct
200+0 records in
200+0 records out
209715200 bytes (210 MB) copied, 0.392785 s, 534 MB/s

--- Additional comment from Steve D on 2015-09-02 16:15:25 EDT ---

Forgot to mention throughput when arbiter node is down inside a VM is very similar to  the fuse mount point:

VM:
dd if=/dev/zero of=test bs=1M count=200 oflag=direct
200+0 records in
200+0 records out
209715200 bytes (210 MB) copied, 2.80659 s, 74.7 MB/s

fuse:
dd if=/dev/zero of=test bs=1M count=200 oflag=direct
200+0 records in
200+0 records out
209715200 bytes (210 MB) copied, 2.55501 s, 82.1 MB/s

--- Additional comment from Vijay Bellur on 2015-09-03 00:28:14 EDT ---

REVIEW: http://review.gluster.org/12095 (afr: Do not wind the full writev payload to arbiter brick) posted (#1) for review on master by Ravishankar N (ravishankar)