On 10/10/15, 11:35 PM, "Clint Byrum"<cl...@fewbar.com> wrote:
Excerpts from Alec Hothan (ahothan)'s message of 2015-10-09 21:19:14
-0700:
On 10/9/15, 6:29 PM, "Clint Byrum"<cl...@fewbar.com> wrote:
Excerpts from Chris Friesen's message of 2015-10-09 17:33:38 -0700:
On 10/09/2015 03:36 PM, Ian Wells wrote:
On 9 October 2015 at 12:50, Chris
Friesen<chris.frie...@windriver.com
<mailto:chris.frie...@windriver.com>> wrote:
Has anybody looked at why 1 instance is too slow and what it
would take to
make 1 scheduler instance work fast enough? This does
not preclude the
use of
concurrency for finer grain tasks in the background.
Currently we pull data on all (!) of the compute nodes out
of the database
via a series of RPC calls, then evaluate the various filters
in python code.
I'll say again: the database seems to me to be the problem here.
Not to
mention, you've just explained that they are in practice holding
all the data in
memory in order to do the work so the benefit we're getting here
is really a
N-to-1-to-M pattern with a DB in the middle (the store-to-DB is
rather
secondary, in fact), and that without incremental updates to the
receivers.
I don't see any reason why you couldn't have an in-memory scheduler.
Currently the database serves as the persistant storage for the
resource usage,
so if we take it out of the picture I imagine you'd want to have
some way of
querying the compute nodes for their current state when the
scheduler first
starts up.
I think the current code uses the fact that objects are remotable
via the
conductor, so changing that to do explicit posts to a known
scheduler topic
would take some work.
Funny enough, I think thats exactly what Josh's "just use Zookeeper"
message is about. Except in memory, it is "in an observable storage
location".
Instead of having the scheduler do all of the compute node inspection
and querying though, you have the nodes push their stats into
something
like Zookeeper or consul, and then have schedulers watch those stats
for changes to keep their in-memory version of the data up to date. So
when you bring a new one online, you don't have to query all the
nodes,
you just scrape the data store, which all of these stores (etcd,
consul,
ZK) are built to support atomically querying and watching at the same
time, so you can have a reasonable expectation of correctness.
Even if you figured out how to make the in-memory scheduler crazy
fast,
There's still value in concurrency for other reasons. No matter how
fast you make the scheduler, you'll be slave to the response time of
a single scheduling request. If you take 1ms to schedule each node
(including just reading the request and pushing out your scheduling
result!) you will never achieve greater than 1000/s. 1ms is way lower
than it's going to take just to shove a tiny message into RabbitMQ or
even 0mq.
That is not what I have seen, measurements that I did or done by
others show between 5000 and 10000 send *per sec* (depending on
mirroring, up to 1KB msg size) using oslo messaging/kombu over
rabbitMQ.
You're quoting througput of RabbitMQ, but how many threads were
involved? An in-memory scheduler that was multi-threaded would need to
implement synchronization at a fairly granular level to use the same
in-memory store, and we're right back to the extreme need for efficient
concurrency in the design, though with much better latency on the
synchronization.
These were single-threaded tests and you're correct that if you had
multiple threads trying to send something you'd have some inefficiency.
However I'd question the likelihood of that happening as it is very
likely that most of the cpu time will be spent outside of oslo
messaging code.
Furthermore, Python does not need multiple threads to go faster. As a
matter of fact, for in-memory operations, it could end up being slower
because of the inherent design of the interpreter (and there are many
independent measurements that have shown it).
And this is unmodified/highly unoptimized oslo messaging code.
If you remove the oslo messaging layer, you get 25000 to 45000
msg/sec with kombu/rabbitMQ (which shows how inefficient is oslo
messaging layer itself)
So I'm pretty sure this is o-k for small clouds, but would be
a disaster for a large, busy cloud.
It all depends on how many sched/sec for the "large busy cloud"...
I think there are two interesting things to discern. Of course, the
exact rate would be great to have as a target, but operational security
and just plain secrecy of business models will probably prevent us from
getting at many of these requirements.
I don't think that is the case. We have no visibility because nobody
has really thought about these numbers. Ops should be ok to provide
some rough requirement numbers if asked (everybody is in the same boat).
The second is the complexity model of scaling. We can just think about
the actual cost benefit of running 1, 3, and more schedulers and come up
with some rough numbers for a lower bounds for scheduler performance
that would make sense.
If, however, you can have 20 schedulers that all take 10ms on average,
and have the occasional lock contention for a resource counter
resulting
in 100ms, now you're at 2000/s minus the lock contention rate. This
strategy would scale better with the number of compute nodes, since
more nodes means more distinct locks, so you can scale out the number
of running servers separate from the number of scheduling requests.
How many compute nodes are we talking about max? How many scheduling
per second is the requirement? And where are we today with the
latest nova scheduler?
My point is that without these numbers we could end up
under-shooting, over-shooting or over-engineering along with the
cost of maintaining that extra complexity over the lifetime of
openstack.
I'll just make up some numbers for the sake of this discussion:
nova scheduler latest can do only 100 sched/sec for 1 instance (I
guess the 10ms average you bring out may not be that unrealistic)
the requirement is a sustained 500 sched/sec worst case with 10K
nodes (that is 5% of 10K and today we can barely launch 100VM/sec
sustained)
Are we going to achieve 5x with just 3 instances which is what most
people deploy? Not likely.
Will using more elaborate distributed infra/DLM like consul/zk/etcd
going to get us to that 500 mark with 3 instances? Maybe but it will
be at the expense of added complexity of the overall solution.
Can we instead optimize nova scheduler with single instance to do
500/sec? Maybe but if we succeed we'll get a lot more simple solution.
Of course we can. And simple solutions are great. So if you can get one
node to do all the work and the cloud doesn't fall over dead when it
dies
(because you have more waiting on standby), that would be fantastic.
I'm dubious that this will be a solution that works for large deployers.
But I may be wrong!
Control plane apps that had to do more complex work than scheduling
instances have been working for a long time at scale using simple
active/standby designs.
In this case it is easier because we can even afford some scheduling
disruption when your active goes down (as long as the standby can pick
up most of the pending requests). Heck when an entire openstack
controller goes down, I bet failing a few request will be the least of
your concerns (because a lot more other things will go wrong).
Not saying that high concurrency and distributed schedulers is not
the solution and maybe we really need a distributed solution, but
it'd be good to have some numbers to frame the discussion.
Indeed, however, I don't think we can expect those numbers to
materialize
in public, ever. We have to make some informed guesses and see how the
product managers respond when we tell them what we think should
happen. ;)
Isn't that a concern? Shouldn't the TC provide at least some numbers
for the scale range so we do not under/over engineer? Any number would
be better than no number at all.
It is an open secret that Nova/Neutron does not scale well below the
1000-node mark (400/500 seems to be a threshold to do anything serious
in production).
If we look at all the openstack/Neutron deployments today, nobody has
an exact count, but very likely the vast majority of deployments are
smaller than 100 nodes. Few are over 100 nodes, even less over 500
(well we know some companies/orgs have deployed thousands but none
with Neutron).
For those 99% of deployments, we only need to service less than 100
nodes. In that space the inability to schedule over 100 instances per
second is probably not that important. If the problem we are trying to
solve here is "scheduling is too slow", then the first question to ask
is how much faster do we need to go, then brainstorm on what would be
the best way to achieve.
I have no problem believing we can schedule a lot more instances/sec
by using scale out design with DLMs but there is no free lunch. Is it
fair to impose on those 99% deployments the complexity of a machinery
that is designed to handle 1000 schedules/sec? Ops are very concerned
about adding even more complexity to an already complex platform and I
think we should be considerate to that, meaning worry about the cost
impact in HW, config, deployment, troubleshooting.
