By default, the client reads from localhost first, if the client is also
a server, or the first to respond. This can be tuned to balance the load
better (see "gluster volume set help") but that's not necessarily
more
efficient. As always, it depends on the workload.
Reading as you suggested is actually far less efficient. The reads would
always be coming from disk and never in any readahead cache.
On 10/03/2016 02:24 AM, Gandalf Corvotempesta wrote:> Hi to all.
> I know that when writing, a client is writing to all replica at once,
> thus the transfer rate is total_bandwidth/replica_count
>
> But what about reads? Client is able to read from multiple nodes at
> once? Which kind of data is reading?
>
> Let me try to explain:
>
> let's assume a 3MB file stored on 3 bricks on 3 servers (1 brick per
> server). For semplicity, let's assume 1MB stored on each brick.
>
> When client reads, will read the first MB from all of 3 servers in
> parallel, then it moves to the secodn MB and so on, or is smart enough
> to read the first MB from server1, the second from server2 and the
> third from server3 at the same time?
>
> I think that second case is way faster than first.
> _______________________________________________
> Gluster-users mailing list
> Gluster-users at gluster.org
> http://www.gluster.org/mailman/listinfo/gluster-users