[openstack-dev] [trove] My thoughts on the Unified Guest Agent

Tim Simpson tim.simpson at rackspace.com
Thu Dec 19 00:57:05 UTC 2013


>> python:
>> consumes 12.5MB of virt memory and 4.3MB of resident memory.

Very few Python only processes will take up such a small amount of virtual memory unless the authors are disciplined about not pulling in any dependencies at all and writing code in a way that isn't necessarily idiomatic. For an example of a project where code is simply written the obvious way, take the Trove reference guest, which uses just shy of 200MB of virtual memory and 39MB of resident. In defense of the reference guest,  there are some things we can do there to make that figure better, I'm certain. I just want to use it as an example of how large a Python process can get when the authors proceed doing things the way they normally would. 

>> C:
>> 4MB of virt memory and 328k of resident memory

>> C++:
>> 12.5MB of virt memory and 784k of resident memory

Much of the space you're seeing is from the C++ standard library. Building a process normally, I get similar results. However it is also possible to statically link the standard libraries and knock off roughly half of that, to 6.42MB virtual and 400kb resident.

Additionally, the C++ standard library can be omitted if necessary. At this point, you might argue that you'd just be writing C code, but even then you'd have the advantages of template metaprogramming and other features not present in plain C. Even without those features, there's no shame in writing C style code assuming you *have* to- C++ was designed to be compatible with C to take advantages of its strengths. The only loss would be some of the C99 stuff like named initializers.

Additionally, in a vast number of contexts the virtual memory "used" for the standard library is not going to matter as other processes will be including that code anyway.

Going back to the Trove C++ Agent, it takes 4MB of resident and 28MB of virtual memory. This is with some fairly non-trivial dependencies, such as Curl, libz, the MySQL and Rabbit client libraries. No special effort was expended making sure we kept the process small as in C++ things naturally stay tiny.

>> C++ is full of fail in a variety of ways and offers no useful advantage for something as small as an agent ;-)

If you haven't recently, I recommend you read up on modern C++. The language, and how it's written and explained, has changed a lot over the past ten years.

Thanks,

Tim


-----Original Message-----
From: Steven Dake [mailto:sdake at redhat.com] 
Sent: Wednesday, December 18, 2013 4:15 PM
To: OpenStack Development Mailing List (not for usage questions)
Subject: Re: [openstack-dev] [trove] My thoughts on the Unified Guest Agent

On 12/18/2013 12:27 PM, Tim Simpson wrote:
>>> Please provide proof of that assumption or at least a general hypothesis that we can test.
> I can't prove that the new agent will be larger as it doesn't exist yet.
>
>>> Since nothing was agreed upon anyway, I don't know how you came to that conclusion.  I would suggest that any agent framework be held to an extremely high standard for footprint for this very reason.
> Sorry, I formed a conclusion based on what I'd read so far. There has been talk to add Salt to this Unified Agent along with several other things. So I think its a valid concern to state that making this thing small is not as high on the list of priorities as adding extra functionality.
>
> The C++ agent is just over 3 megabytes of real memory and takes up less than 30 megabytes  of virtual memory. I don't think an agent has to be *that* small. However it won't get near that small unless making it tiny is made a priority, and I'm skeptical that's possible while also deciding an agent will be capable of interacting with all major OpenStack projects as well as Salt.
>
>>> Nobody has suggested writing an agent that does everything.
> Steven Dake just said:
>
> "A unified agent addresses the downstream viewpoint well, which is 'There is only one agent to package and maintain, and it supports all the integrated OpenStack Program projects'."
>
> So it sounds like some people are saying there will only be one. Or that it is at least an idea.
>
>>> If Trove's communication method is in fact superior to all others, then perhaps we should discuss using that in the unified agent framework.
> My point is every project should communicate to an agent in its own interface, which can be swapped out for whatever implementations people need.
>
>>>   In fact I've specifically been arguing to keep it focused on facilitating guest<->service communication and limiting its in-guest capabilities to narrowly focused tasks.
> I like this idea better than creating one agent to rule them all, but I would like to avoid forcing a single method of communicating between agents.
>
>>> Also I'd certainly be interested in hearing about whether or not you think the C++ agent could made generic enough for any project to use.
> I certainly believe much of the code could be reused for other projects. Right now it communicates over RabbitMQ, Oslo RPC style, so I'm not sure how much it will fall in line with what the Unified Agent group wants. However, I would love to talk more about this. So far my experience has been that no one wants to pursue using / developing an agent that was written in C++.
>
> Thanks,
>
> Tim
> ________________________________________
> From: Clint Byrum [clint at fewbar.com]
> Sent: Wednesday, December 18, 2013 11:36 AM
> To: openstack-dev
> Subject: Re: [openstack-dev] [trove] My thoughts on the Unified Guest 
> Agent
>
> Excerpts from Tim Simpson's message of 2013-12-18 07:34:14 -0800:
>> I've been following the Unified Agent mailing list thread for awhile 
>> now and, as someone who has written a fair amount of code for both of 
>> the two existing Trove agents, thought I should give my opinion about 
>> it. I like the idea of a unified agent, but believe that forcing 
>> Trove to adopt this agent for use as its by default will stifle 
>> innovation and harm the project.
>>
> "Them's fightin words". ;)
>
> That is a very strong position to take. So I am going to hold your 
> statements of facts and assumptions to a very high standard below.
>
>> There are reasons Trove has more than one agent currently. While 
>> everyone knows about the "Reference Agent" written in Python, 
>> Rackspace uses a different agent written in C++ because it takes up 
>> less memory. The concerns which led to the C++ agent would not be 
>> addressed by a unified agent, which if anything would be larger than 
>> the Reference Agent is currently.
>>
> "Would be larger..." - Please provide proof of that assumption or at 
> least a general hypothesis that we can test. Since nothing was agreed 
> upon anyway, I don't know how you came to that conclusion. I would 
> suggest that any agent framework be held to an extremely high standard 
> for footprint for this very reason.
>
>> I also believe a unified agent represents the wrong approach 
>> philosophically. An agent by design needs to be lightweight, capable 
>> of doing exactly what it needs to and no more. This is especially 
>> true for a project like Trove whose goal is to not to provide overly 
>> general PAAS capabilities but simply installation and maintenance of 
>> different datastores. Currently, the Trove daemons handle most logic 
>> and leave the agents themselves to do relatively little. This takes 
>> some effort as many of the first iterations of Trove features have 
>> too much logic put into the guest agents. However through 
>> perseverance the subsequent designs are usually cleaner and simpler 
>> to follow. A community approved, "do everything" agent would endorse 
>> the wrong balance and lead to developers piling up logic on the guest 
>> side. Over time, features would become dependent on the Unified 
>> Agent, making it impossible to run or even contemplate light-weight agents.
>>
> Nobody has suggested writing an agent that does everything. A 
> framework for agents to build on is what has been suggested. In fact 
> I've specifically been arguing to keep it focused on facilitating 
> guest<->service communication and limiting its in-guest capabilities 
> to narrowly focused tasks.
>
>> Trove's interface to agents today is fairly loose and could stand to 
>> be made stricter. However, it is flexible and works well enough. 
>> Essentially, the duck typed interface of the trove.guestagent.api.API 
>> class is used to send messages, and Trove conductor is used to 
>> receive them at which point it updates the database. Because both of 
>> these components can be swapped out if necessary, the code could 
>> support the Unified Agent when it appears as well as future agents.
>>
>> It would be a mistake however to alter Trove's standard method of 
>> communication to please the new Unified Agent. In general, we should 
>> try to keep Trove speaking to guest agents in Trove's terms alone to 
>> prevent bloat.
>>
> If Trove's communication method is in fact superior to all others, 
> then perhaps we should discuss using that in the unified agent framework.
>
> Also I'd certainly be interested in hearing about whether or not you 
> think the C++ agent could made generic enough for any project to use.
> That would be a nice win.
If footprint is the concern, I am certain you would be better off with a C implementation.  C++ is full of fail in a variety of ways and offers no useful advantage for something as small as an agent ;-)

We wrote our heat agents in python because they are not long-lived processes, hence they can bleed into the guest OS memory on demand and are only typically used during bootstrapping, vs always taking a chunk off the top as a long-lived agent process would do.

Just for comparison sake, the basic footprint of an operation like
sleep(100) (eg the runtime)

python:
consumes 12.5MB of virt memory and 4.3MB of resident memory.

C:
4MB of virt memory and 328k of resident memory

C++:
12.5MB of virt memory and 784k of resident memory

C++ buys 3MB of ram, compared to all the losses of the great
functionality the infrastructure team has put into place.  But C is a real winner for footprint, especially since the C library (the virt
part) will almost always be demanded by some application via the dynamic loader.

Regards
-steve

> _______________________________________________
> OpenStack-dev mailing list
> OpenStack-dev at lists.openstack.org
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>
> _______________________________________________
> OpenStack-dev mailing list
> OpenStack-dev at lists.openstack.org
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev


_______________________________________________
OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev



More information about the OpenStack-dev mailing list