Sustainable Computing at Scale for Smart Cities

Sustainable Computing at Scale for Smart Cities


Good afternoon. So I’m Yogesh, I’m with the CDS
department here at IISc. And I’m a researcher by training
I’m not an AI researcher. So I’ll slightly take a liberal
view of sustainability not necessarily sustainability for
the cities. But it’s a two step thing to
fascinated would be of cities. I’ll be talking about how
we can, I’ll be drawing on experiences from a few projects
I’ve seen involved with. Some actually starting back
at USC as part of my earlier overture, as of such faculty. But also here at IACS, part
of a smart campus project for water management, and also as
part of a smart city testbed, which Baruraj will be
talking about next. But broadly, this gets into the
area of being able to collect data as part of the IoT,
the internet of things. Installation of sensors and
actuators through various city and community
scales [INAUDIBLE] right. [INAUDIBLE] mentioned about
how you can use sensing at the classroom level,
and by building models for doing it within the classroom,
so. You’ve done some work on smart
water management to bring water level sensors in water tanks and
so on. And it’s about smarter desk we
have things like cameras and so on going off the field. But a key point that we hear is
that you have the ability to measure a lot of parameters in
an automated fashion between session throughout our
urban environments, right? And the question really comes
into what you can do with such data, right? And typically, if you look
at analytics broadly, and AI is just one part of
an analytics pipeline. It’s about acquiring the data,
cleaning it, integrating it, storing it and doing online
analytics on top of that, building models on that,
and so on. And finally closing the loop
where you actually go and perform an intervention, right. And disclosing the loop
has served two aspects, one is conceptually
closing the loop, right. I measure let’s say temperature
in the room, and you based on that and the current
temperature in the models and so on, I’m gonna go and
intervene on the AC unit right. But it’s also a physical
closing the loop because often times what happens is that
this acquisition pipeline might go through a large computational
cycle, but typically the sensing and the activation of physically
approximate to each other. And that’s a important point. The key point is that itself
is enabled by the ability to collect data and
availability of all this data. And AI is going to help with
the sustainable and so on. So one of the key benefits of
all these data being out there is sort of the advances we’ve
seen in deep learning, right? So you’ve got a lot of data
collected from all around, and you have also. So access to computing resources
unlike anything before. You can sort of on the top
of a pen you can go and get computers on
your favorite cloud. Or buy a GPU and do way more
than you could a few years back. And this context we soon
realized we needed, especially looking at
scalable systems and so on, a lot of the data collected
from census, like temperature, and water and so on. It’s not large in size, and
they’re typically point observations about a specific
type of entity, right? When we started working
with video data, we didn’t realize that video
is actually a meta-sensor. You can actually do a lot more
with just a single frame of video than I could with maybe
20 sensors I install over there. Partly because you have context, you actually have a bunch
of different parameters. So video data combined
with deep learning can actually be a very
powerful mechanism. And that essentially tells you
that one form of sustainability is actually reducing
how much computing devices we put out there. Or how much physical devices
you put out there, right? So we’re looking at how
can you actually reduce. The physical deployment and do more with video streams and
AI, right? And in this context, taking you
to the computing side of things, we realize as part of various
city and campus deployments, that we are not just deploying
sensors and activators. But we are actually deploying
computing devices in the field without realizing it up front. Like Raspberry Pi has actually got more
computing power way more. Than maybe a PC had
ten years back and often times these are used
just to collect it and push them off of the cloud
before the posting, right? And we ran a bunch of
experiments in terms of the computer capacity of these
Raspberry Pi and so on, and we actually found there are almost,
three Raspberry Pi are almost the equal of, let’s say, one
virtual machine on the cloud. So one of the advantages
of Cloud computing, which typically is used in many
of the smart city projects, is that you have elastic
available different sources you can acquire and least machines. You’re gonna have
to maintain them so these are some of
the good parts. But some of the not so good
parts are the facts that you have a lot more network latency
going from the edge where the physical infrastructure is
to the cloud in the data center. And a bandwidth can sometimes be
constrained if you think about how bandwidth applications
like video, right. So we started looking at how
can you make use of these edge computing devices and
part of let’s say RF pi devices. And you are starting to
see even GPU enabled edge devices sometimes
called Fog Devices. Like in [INAUDIBLE] that costs
maybe a couple hundred dollars. Which could actually be on the
field close to where the date is being collected. These are good in a sense that
the network latency between the data collection,
decision making and the control can be reduced and then bonded potentially within
a smaller private network is much more available than
between the Edge and the Cloud. But you also have challenges
such as it’s not as elastic, in fact it is not elastic
compare to let’s say, Cloud which of machines. And there are reduced
capacities, and battery availability and
reliability and so on, that you have to deal with. But for us, the broad question
was if you combine Edge devices with Cloud devices, can you do more than just
using Cloud resources? And one of the big challenges
that we notice when it came to using such a Edge devices for
making decision making within smart city is there is
no easy way to program them or access them or deploy
applications on them right? That’s sort of like to,
small activity, right. We looked at how you can take,
just like you would, let’s say, compose a spark or a or a job. And you would run them and let the platform figure out
where things run and so on. Can you do that more
broadly across Edge and Cloud devices, right? So here we essentially put
together a small platform, which we call ECHO not
related to Amazon’s ECHO. And so unfortunate
choice of Acronym but the idea here is that you can
compose your own decision-making application as a bag of tasks. We actually support both
streaming data, which is typical for sensor-based data as well as
batch data which might be more common for doing training
on historic data and so on. And all of the things that we
tried not to do with claimed to come up with the next big
Hadoop or Spark or so on. Basically said that, hey, you
said you want a thin layer that can allow you to access
various distributed platforms. But, make use of existing
software platforms that already exist, right? So things like Edgent,
or Storm, or Nifi, or so on, which already are doing,
in the Big Data Platform is doing a very insort
job we plug into those. This will give us a facade
on top of them and we enable, one of the key
advantages of enable is orchestration across various
types of these Edge, Fog and Cloud devices in
a disputed setup. And a couple of interesting
things that we’re sort of recognisant of when we
started actually by sort of trying to apply it and
so on. Is that when we have
interrupting IoT deployment, there’s a lot network as
symmetry in the sense that a lot of sensors and might be
behind private networks. You might be able to connect
from one device to another but not the other way around. So the of our deployment and orchestration of
these applications. We make sure that
there is connectivity, in either one of
the two directions or it goes through things like
third parties and so on. These are things that you
don’t want the developer to think about and
how the platform, give you out of the box,
is something that we do. Another thing to be cognizant
of is the fact that you have devices that could be
dying once you put them on the field, right? So trust me, when it comes
to hardware deployment and maintaining it on the field
it is not easy, right? As software [INAUDIBLE] we
actually have life easy. And really pretty people who
walk on the field with hardware devices. So you actually want to make
sure that if things go kaput, you can actually move
things on the fly, right? Migrate them from less
applications that are running on one device, often on other
device that’s available, based on the current situation
that’s out there, right? Next, I will give you sort of
a high level overview of how things work you have a bunch
of different devices. Edge, Fog, Cloud is where we
have our platform services that central co-platform
services run on the Cloud. But also bunch of really
lightweight services that run on various computing devices
which you might be using. And on top of that we also
allow users to bring in their own favorite runtime
engine be it Nifi intense flow, Engine storm and so on. And then when
a user comes in and wants to run that application,
right. Say they might be doing some
predictive modeling and so on. You can sort of bring it in and
that gets automatically deployed across the various devices based
on what’s available and so on. And the whole thing gets going,
and you can rebalance it and things like that,
as might be the need, right? Just to give you a flavor of
some of the applications that we have got working on
this ECHO platform, and all of these actually are real
applications, based on our experiences with our smart
city and smart projects. So one is just a classic ET
piper, where you want a store into and later on come back and
take a look at it. Here we have the original
version using Apache storm as a stream processing on
the cloud exclusively. But with a ECHO we can do
part of the reprocessing and the parsing and
the data transmission as one. On the Edge device instead of
just handing it off to Cloud, once they go past here, right? Another slightly more
interesting application, especially in
the context of AI and deep learning, is when we’re
actually doing video analytics, when we have streams of
video data coming in and we want to do classification,
right? And here you might actually
want to do low-granularity classification on the edge, and if something interesting comes
out, you might actually want to push it off to the cloud
with a high NGPU and then do more concrete so
you spend a bandwidth only if something interesting comes
between the agent and the Cloud but otherwise do
things in the edge, right? So these are a couple of
examples, we have plenty of examples as well but if you want
you can actually play it out. You can download this
echo platform and play around with it,
it’s available on our page. Just to sort of conclude bunch
of interesting possibilities that come out, this is more
a platform play so far but there’s much more interesting
research coming out of that. How do you compose the DAGs? So that users compose it and the system automatically
figures out what engine to use. Right now it’s
really prescriptive, the user specifies and
do it automatically. How can you schedule in
an automated fashion. And down the line how can you
even do distributed training on some of these deep
learning models? Because that oftentimes is
the biggest bottleneck. Right now we focus more
on the inferencing part. So, again, bunch of
collaborators to thank and I thank you all for
your attention.>>[APPLAUSE]
Quick question?>>Yeah.>>Yeah.>>[INAUDIBLE]
>>We look at a pipe class divisor higher, so depends
on, for example, 1GB memory and one gigahertz CPU core,
might be the baseline, right? But it won’t work
with let’s say, low end,
it’s not designed for that.

Leave a Reply

Your email address will not be published. Required fields are marked *