Ros 6.0rc6 x86


I have 30 KVM's running
I run traffic generator to the bridge interface that they are all connected to just bench testing, I decided to enable a que on each interface to test, how the system handles

20mb/s traffic gen to the bridge interface, and apply a 10m/s QUE to each interface... START

Issue is:
queue from 25 - 31 Fail
Please read carefully.
at first I start and only queue 1-24 show any stats on the queue tree -HOWEVER the interface is actualy queueing corrrectly at the set speed (10m/s)
issue is it wont show any stats at all

so i decided to disable queue 1-24 and enable only 25-31

then and only then did any stats show.

See picture below is when i -REENABLE- all queue's, and then Stop the traffic generator
as you can see 25-31 are stuck at 9.9mb/s but there is absolutly no traffic on any interface.

Also why is the management of the KVM so problematic


starting stopping the kvm is not good, some times it dont stop and get stuck on "starting" or "restarting"
half the time its still running, and what makes it worse when its in this stuck state, you can no longer control it, most of the time the console wont connect

all you are left with is a forced reboot of the host machine to actualy get the kvm to start again

That or Remove the kvm and re add it. it makes it pretty much impossible to make it a production machine...

staff?

Its almost as if routeros cant handle too many queues with alot of traffic ?


is there anyone out there with an insane amount of queues and has alot of bandwidth going over them?

is this an x86 issue ? or a routeros issue?

I personally am going for its a slight "lag" issue perhaps there are measures in routeros to not use too much resources for display data?

while screenshot shows awesome resolution of your screen it does not tell exactly what is happening.

First of - if you are running something in KVM, you can do just the same on host system, to see if that is load issue or something different since packets for guest are manged by host, then by guest and then by host again (if traffic is passing through the guest)

Also, where you are generating the traffic? on the same host adding even more load or somewhere else?

What is purpose of this test?

as different configurations have different possible implementations, while you can have 1000 clients and can have 1000 simple queues for each of them, you also can do the same just having queues for user groups, like starter pack queue, normal queue and high bandwidth queue and all of them using PCQ to manage load. And there will be huge difference in latency having 1000 or 3 queues in configuration.

And there will be huge difference in latency having 1000 or 3 queues in configuration.

on the presentation, haven't Janis said that v6 has optimized simple queues with almost no differences compared to pcq?.. %) I can be wrong...

while screenshot shows awesome resolution of your screen it does not tell exactly what is happening.

First of - if you are running something in KVM, you can do just the same on host system, to see if that is load issue or something different since packets for guest are manged by host, then by guest and then by host again (if traffic is passing through the guest)

Also, where you are generating the traffic? on the same host adding even more load or somewhere else?

What is purpose of this test?

as different configurations have different possible implementations, while you can have 1000 clients and can have 1000 simple queues for each of them, you also can do the same just having queues for user groups, like starter pack queue, normal queue and high bandwidth queue and all of them using PCQ to manage load. And there will be huge difference in latency having 1000 or 3 queues in configuration.

I too thought they were optomised in V6

I will explain this, i apologize it was not so clear
At this moment all of the kvm's are not in production, they have absolutly no config other than an IP address assigned from the Host machine, they don't even have any queues, i'm explaining this in case its assumed that we want to see the load of each individual KVM, as there might be a separate load there via other reasons.

I could provide more screenshots, but i can assure you that each KVM is not 100% with in it self.
How ever the Host System has the cpu usage very high, because well from what i can understand
the traffic generator which is on the Host system is generating the packets, and sending to its own local bridge, which is the bridge that has the subnet containing the KVM's.

I simply targeted the broadcast domain, and the KVM's are showing the traffic response to the Host machine perfectly on the interface menu.

Queue tree's 1 for each "virtual Ethernet Interface" the queues respond quite clearly
But with Simple Queues, it has very strange behavior, which is on display, if you look closely, you can see that there is strangely
Some queues that have a high amount of traffic at that "instance", but the issue is, im not even sending any data, when i took that screenshot.

So if i was to cut to the chase of the issue, the last say 10 or so queues are stuck in motion the numbers do not change, UNLESS
i fiddle around and disable and enable the say 0-20 on and off, then ! and only then ! do the last 10 or so queues the figures/numbers change very quickly.

I do hope that has explained it more, the purpose of my test was, to bench test a Dell server, i want to test it stability, research an development you could say, I want to provide this as a solution in out data center, and this is the makings to see of this is up to standard and stable.

if your end result will use KVM guests, use them, if not, then doing tests that include KVM guest is useless due to this:
1) packet is generated on the router - this load will never be on the router
2) packet is passed to vif via bridge or routing
3) packet is processed by guest received on vif, and then sent back via vif
4) packet is received by host

this is quite a lot of overload if your end configuration will not use KVM guest at all. If you will have guest, you have to test this and have to have guest configuration in place.

about simple queues - in ROS 6.0 simple queues have very little impact, that is ok, just make sure that each queue matches its own traffic. we have seen no issues so far in testing.

so, preferably, move traffic generator elsewhere, something like iperf is fine too, then pass all that traffic into the router and see what happens when you hit the simple queues.

janisk although this is old I cant seem to get a response via support

The tests are firstly, okay lets load a ton of kvm VM's give it a good thrashing
and lets see how much traffic i can pump though it, I will then decide ok this system really can only handle X amount
that is all , perhaps my target for the traffic generator was too agressive and should be routed or something

but as i mentioned before about the queues, even though you have not seen that issue, you have seen the issue in this thread, so I hope its helpfull for you

janisk im also trying to get a response to more testing we did via our emails

thanks