Requirements of computing
and network joint optimization and scheduling
China Mobile
No.32 XuanWuMen West Street
Beijing
100053
China
fuyuexia@chinamobile.com
China Mobile
No.32 XuanWuMen West Street
Beijing
100053
China
liupengyjy@chinamobile.com
China Mobile
No.32 XuanWuMen West Street
Beijing
100053
China
gengliang@chinamobile.com
Computing in Network Research Group
computing, joint optimization
With the development of edge computing, there is a trend that
computing is widely deployed in network rather than at other end of
network, and provides services at nearer location. With the deep
integration of network, traditional optimization and scheduling within
network domain is not enough, the endpoint of the path matters a lot. So
the relationship between computing and network are new and important
topics to be studied. This document focus on the requirements of
computing and network joint optimization and scheduling based on the
newly arising service requirements.
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
document are to be interpreted as described in .
For traditional services without strict service requirements, the
best-effort network can meet the requirements with traditional path
optimization, which only consider the network condition. With new
services arising, such as cloud AR/VR, cloud gaming, V2X, new and strict
requirements towards network, also towards the service endpoint are
proposed to meet the service requirements. So the computing and network
joint optimization and scheduling are proposed to guarantee the service
performance.
The computing and network joint optimization means that there is not
only the path optimization in network, but also the endpoint joint
optimization; also the two will affect each other. Based on the joint
optimization, the service scheduling can be performed considering the
network condition and also the endpoint condition, with the
“optimal path+ optimal endpoint” policy. What’s more,
the computing and network resources joint reservation is required for
services with strict performance requirements.
The service requirements arising include both network and computing
requirements, which further require future network should perform the
joint optimization according to service requirements. So the
requirements towards the joint optimization are: the awareness of
computing and network requirements, the awareness of computing resources
and services in network, the computing-aware path optimization and the
network-aware endpoint optimization.
2.1. Awareness of computing and network requirements
Awareness of computing and network requirements refers to consider
the computing requirements in addition to the network requirements,
including the awareness of computing requirements and the measurement of
computing services.
Since network requirements can be measured with bandwidth, delay etc,
it is also required to measure computing requirements in a unified way.
On the one hand, there are many different computing services which are
the “consumers” of computing resources, such as video
processing, image classification etc, and they propose various
requirements towards computing. It is required to firstly obtain the
computing requirements and then model the requirements in a unified way,
which then can be used as the constraint of joint optimization.
What’s more, the computing service modes are abundant compared
to network services, which are the computing “producers”,
including there are heterogeneous hardware such as GPU, CPU, FPGA etc,
and also various algorithms deployed in network, so it is also required
to model the computing producers in a unified way, which is another
important factor for joint optimization. As for the awareness of
computing requirements, some technologies such as application-aware
networking have proposed corresponding technical solutions to delivery
computing requirements in the packet head, however, it needs further
study on the security of application and also the efficiency of the
information delivery. As for the measurement of computing services,
there is no mature solution to model the computing requirements and the
computing resources in a unified way, which is a challenge for the
computing and network joint optimization.
2.2. Awareness of computing resources and services
With the development of edge computing, the computing resources and
computing services will be distributed in network, since the limited
physical conditions, each computing site will be small scale and with
limited computing resources, so different from the cloud computing,
which can finish the computing task within one site, the edge computing
sites need the collaboration among many sites, and this collaboration
can be done in network. To coordinate the computing sites, it is
required for network to be aware of the computing status of edge sites,
including the real-time status of computing resources and computing
services. So how to generate the required information and then broadcast
it to network brings new challenges.
3.1 Computing-aware path optimization
With new services requiring computing and network resources,
traditional network-based path optimization can not accurately guarantee
the service requirements. The network-based path optimization only
according to network conditions can only make sure the performance of
network services, it can only find a best path towards a given endpoint,
however, the given endpoint may be not optimal, causing the service
requirements cannot be met.
So It is required to do the computing-aware path optimization to
consider the status of endpoint. For example, before the path
optimization, according to the awareness of computing resources and
services in network, including the location and status, the network
could firstly find a list of optimal computing nodes, then the network
could do path optimization with different computing endpoints, which
changes the traditional way to only do the path optimization with one
destination.
To better optimize the computing-aware path, we need to consider
different weights of computing and network metrics when calculating the
optimal path. For traditional path optimization, there are only network
metrics as the parameters of algorithm; it is required to add computing
metrics also as the calculation metrics of the algorithm and to combine
the computing and network metrics.
What’s more, based on the awareness of service requirements,
for different services, there will be different requirements towards
computing and network. For some computing-intensive services, computing
counts more on the whole process of services, so they will require more
on computing than network; and for communication-intensive services, the
computing is less during the service process, while there will be
frequent communication, which will propose higher requirement towards
network than computing. So it can be inferred that computing and network
matters differently during the service process for various services.
Based on what discussed above, it is required to adaptively define
different weights of computing and network metrics for different
services, adapting to various service requirements. For example, for the
computing-intensive services, it is required to put more weights on
computing metrics than network metrics, which could be based on the
percentage of predicted computing time in whole time; as for
communication-intensive services, more weights could be put to adapt to
the service requirements.
3.2 Network-aware endpoint optimization
Based on the computing-aware path optimization, there will be the
optimal “path + endpoint” pair, combing the computing and
network status. But there will also be inner scheduling in computing
node, which may also influence the computing time. With proper task
assignment, the computing time could be less to make sure that endpoint
provides the promised services. So it is also required for endpoint to
know the service requirements precisely, otherwise the endpoint will
just do the usual scheduling without considering the service
requirements.
With the network-awareness, the endpoint will know the performance of
network, such as the endpoint will know the transmission time in network
and then calculate the rest of required time, and then it will do the
inner scheduling accordingly.
For services with strict computing requirements, the resource
reservation should include network reservation and computing
reservation, also, the two will affect each other.
There is network resource reservation in traditional QoS guarantee
mechanism based on the network resources reservation calculation to
reserve specific resources for specific services. With new services
arising, the network resources reservation is not enough, since the
completion of services include not only network transmission but also
endpoint calculation, only reserving the network resources cannot make
sure the required computing resources are available during the required
time for specific service.
So facing the trend of computing and network convergence, it is also
required to reserve the computing resources together with the network
resources. Based on the awareness of service requirements and the joint
path optimization, it is required to map the computing requirements into
the corresponding computing resources reservation, for example, to map
the services type into the computing resources type, and translate the
computing latency requirements towards the required amount computing
resources.
On the other hand, the reservation of network and computing resources
are closely linked, there will be different network resource reservation
policy considering the computing resources reservation. For example, the
order of the two resources reservation requires to be considered since
they are relative independent usually.
What’s more, it is also required to dynamically adjust the
resources reservation according to real-time status. One scenario is
that the computing resource reservation could be adjusted based on the
information from network domain, including the reservation time and also
the reservation amount. Another scenario is the co-adjust of the two
resources reservation, in network domain, the path and the relative
reservation could be adjusted, and then the computing domain is required
to adjust on-demand.
Based on the new services’ requirements on computing and
network, this document puts forward requirements of computing and
network joint optimization, and also proposes requirements of computing
and network joint resource reservation. Computing in network is a new
direction, how to collaborate computing and network need further
study.