Hi
Garrett D'Amore wrote:
Luke Schwab wrote:
Hi,
I have recently been running performance tests with zones on several
Sun servers. (280's and 880's), 2 CPU and 8CPUs respectively.
On my 280 servers (2CPU machines), I installed and booted 3 zones and
I get poor performance on across the network (400Mbps). I only
transfer data at 80% of what I can in the global zone with no zones
installed or booted (where I measured 550Mbps). That is a 20%
difference in performance. I tested with the opensource tool called
IPERF.
I originally thought that zones were causing my problems so I decided
to move to another type of hardware, a V880 (8 CPU) machine. When I
ran my performance tests again with 3 zones installed and booted I saw
<1% overhead in the network compared to running in the global zone
with no zones installed or booted. Both of the 880 tests ran around
860 Mbps throughput with less then 1% difference between several
scripted tests.
Can anyone explain why I see this? Both servers have gigabit NICs
installed. The only thing I can think of is that the backplane on the
servers in a bottleneck for the small 280 servers. But how could this
(or something else in the kernel space) be causing the symptoms I see
above?
Any comments would be helpfull,
Basically, running a Zone networking takes CPU. If you have the CPU
resources to spare, then it won't matter much. But if you are CPU
limited, then the impact of a Zone will be more noticeable.
My own experience, and I have re-validated it, is that on a
system with
sufficient CPU/memory, there is such a small difference between
running
three FTP sessions to the global zone and one each to three
non-global
zones that I attribute it to statistical noise. (this was with
shared IP)
I was moving close to 3Gbps across three e1000g interfaces (pretty
much at wire speed). x4100 to x4200 (or vise versa).
However, depending on which build of NV this is, non-global zones now
run java as part of webconsole, and if the system is short on
memory,
running zones could introduce contention. svcadm disable webconsole
might help.
Steffen
One way to get better performance in your zone, I think, may be to use
exclusive IP instances. This requires each zone have its own NIC, but
the benefit is that less CPU is required to forward packets to the
"correct" zone. A compromise (halfway) may be to use VLANs, and put
each zone in its own VLAN, and still use exclusive IP instances.
-- Garrett
Thanks,
ljs
This message posted from opensolaris.org
_______________________________________________
networking-discuss mailing list
[email protected]
_______________________________________________
networking-discuss mailing list
[email protected]
_______________________________________________
networking-discuss mailing list
[email protected]
