Gluster users - Sep 2017 - investigate & troubleshoot speed bottleneck(s) - how?

hi guys/gals

I realize that this question must have been asked before, I 
sroogled and found some posts on the web on how to 
tweak/tune gluster, however..

What I hope is that some experts and/or devel could write a 
bit more, maybe compose a doc on - How to investigate and 
trouble gluster's speed-performance bottleneck.

Why I think such a thorough guide would be important? Well.. 
I guess many of us wonder when we look at how "raw" fs vs 
glusterfs does then we wonder - how come!?
I know such a comparison is oversimplification or maybe even 
unfair but when I see such a gigantic performance difference 
then I think, I hope - it must be possible to help with a 
detective work to unravel any bottlenecks that may hamper 
glusterfs so badly that almost to a point that one 
wonders... what is the point.

I did today such a oversimplified test, I used dbench on a 
raw xfs lvm raid0 four ssd pvs (no hardware raid)

$ dbench -t 60 10
...
8 of 10 processes prepared for launch?? 0 sec
10 of 10 processes prepared for launch?? 0 sec
releasing clients
 ? 10???? 21573? 1339.94 MB/sec? warmup?? 1 sec? latency 
35.500 ms
 ? 10???? 50505? 1448.58 MB/sec? warmup?? 2 sec? latency 
10.027 ms
 ? 10???? 78424? 1467.54 MB/sec? warmup?? 3 sec? latency 
8.810 ms
 ? 10??? 105338? 1462.94 MB/sec? warmup?? 4 sec? latency 
19.670 ms
 ? 10??? 134820? 1488.04 MB/sec? warmup?? 5 sec? latency 
11.237 ms
 ? 10??? 164380? 1505.12 MB/sec? warmup?? 6 sec? latency 
4.007 ms
...
Throughput 1662.91 MB/sec? 10 clients? 10 procs? 
max_latency=38.879 ms

Cluster hosts 9 vols, each vol on three peers in replica 
mode. Generally, but also most of the time vols utilization 
is really low, data would be regular office work, read 
randomly single files.
Peers are connected via net switch stack, each peer connects 
to stack via two-port lacp 1GB link, jumbo mtu.

So I do do same dbench on one of the vols:
...
8 of 10 processes prepared for launch?? 0 sec
10 of 10 processes prepared for launch?? 0 sec
releasing clients
 ? 10??????? 98??? 45.41 MB/sec? warmup?? 1 sec? latency 
113.146 ms
 ? 10?????? 212??? 41.52 MB/sec? warmup?? 2 sec? latency 
93.800 ms
 ? 10?????? 343??? 41.23 MB/sec? warmup?? 3 sec? latency 
53.545 ms
 ? 10?????? 468??? 41.06 MB/sec? warmup?? 4 sec? latency 
54.450 ms
 ? 10?????? 612??? 41.89 MB/sec? warmup?? 5 sec? latency 
152.659 ms
 ? 10?????? 866??? 35.99 MB/sec? warmup?? 6 sec? latency 
31.377 ms
 ? 10????? 1074??? 32.74 MB/sec? warmup?? 7 sec? latency 
39.923 ms
 ? 10????? 1307??? 29.77 MB/sec? warmup?? 8 sec? latency 
42.388 ms
...
Throughput 15.3757 MB/sec? 10 clients? 10 procs? 
max_latency=54.371 ms

So yes.. gee.. how can I make my gluster-cluster faster???

many thanks, L.