You are on page 1of 5

Hyperautomated AI-based

Security vs VMware NSX


Manuel Nedbal

Those tasked with building network security appliances know that network traffic processing is
accomplished in several different stages. Each NextGen Firewall, IPS, and DLP appliance sends traffic
through a pipeline that becomes progressively deeper in their packet analysis. As such, each stage requires
a different amount of compute, storage and memory, and appliance vendors over the years have mastered
how to anticipate required budgets for typical traffic flows. Although the following example oversimplifies
such a processing sequence, and each stage conveys just a glimpse of the complexities involved, we will use
these as a baseline and apply them to our evaluation of the VMware NSX architecture.

1. Packet reception/transmission: picking up and distributing incoming packets to a set of subsequent


processing stages. Receiving outgoing packets and putting them back onto the network

2. Flow processing: reassembling, reordering, re-fragmenting packets to present a stream to


subsequent layers

3. TLS decryption: decrypting traffic and re-encrypting it outbound

4. Deep packet inspection: parsing protocols and delivering the intended appliance features

5. Object handling: extracting files, domains, URLs, etc. and analyzing each

6. Machine Learning: behavioral analysis and deriving anomalies

7. Event processing and correlation: deduplicating, suppressing events, and creating more meaningful
outcomes for security teams

Such security richness is required for detecting VMware has been advocating against monolithic
and countering modern threats. Industry-leading appliances as the security solution for dynamic data
solutions can provide highly effective security if they centers, and they have made significant inroads in
are placed and configured properly. However, as I this direction. Their solution suggests coupling an
noted in the last blog, “Hyperautomation of AI-based ever-richer set of security controls to the hypervisor.
Security in the Distributed Cloud,” security teams Let’s review these solution benefits and cautions,
and then contrast it with a system that supports
are not always able to achieve the required security
Hyperautomated AI-based Security.
effectiveness, and this eventuality leads to
a different set of challenges. For this blog, we VMware’s Networking and Security offering, named
will focus on how VMware’s traffic processing NSX, comes in two flavors, NSX-V and NSX-T. NSX-V
approaches fare in its delivery of rich security is the VMware product that allows traditional
functionality and requirements. appliance vendors to integrate security via an API
HYPERAUTOMATED AI-BASED SECURITY VS VMWARE NSX | 2

called NetX. Via NetX, virtual appliances that run This solution delivers an architecture that scales
on VMware’s hypervisor, can apply traffic inspection out with the number of physical hosts in a data
and control (see Figure 1). In parallel, Access Control center. Rather than aggregating traffic at an
Lists (ACLs) can be enforced inside the VMware artificially created choke point in the network,
hypervisor, which makes sense when considering traffic is inspected close to where it is generated,
the deterministic lookup times required per packet and security capacities keep growing at the same
and the lean processing overhead. rate as the data center.

NSX-V WITH TRADITIONAL SECURITY

Host 1 Host 2
3rd party 3rd party
VM1 VM2 VM3 Security VM VM4 VM5 Security VM

NetX NetX
pg 10 pg 11 pg 10 pg 11
VLAN 10 VLAN 11 VLAN 10 VLAN 11 Virtual Switch

VMware Hypervisor ACLs VMware Hypervisor ACLs

Physical Network

Figure 1: NSX-V with Traditional Security Appliance

While NSX-V provides a major leap in the right •Scaling to the right number of security appliance
direction, there are few issues still left to address: instances depending on traffic flows and policies
does not exist. This starts to become an issue when
•Third-party security VMs are monolithic in nature,
a single security appliance cannot handle the
thus putting a burden of compute, storage and
traffic of all the VMs it needs to protect on the same
memory on each host (all “stages” stick together)
host. The more “steps” need to be processed, and
•Customers must manage a growing set of virtual the more cores Intel adds to a CPU, and the more
appliances from a third-party vendor and make bandwidth can be run over networks, the more
sure they are supporting the VMware versions they likely a bottleneck occurs
want to deploy (there is tight coupling between an
•Upgrades and appliance failures will incur outages
NSX-V agent needed on each security appliance
as there is no High Availability within the same host
and the hypervisor version)

•Chaining multiple monolithic services increases the


latency and overhead problems as the same traffic
processing stages are repeated
HYPERAUTOMATED AI-BASED SECURITY VS VMWARE NSX | 3

Several years after the VMware acquisition of Nicira Given that such an approach requires support from
in 2012, the release of NSX-T hit the market with the the hypervisor, VMware’s technology stack had to
promise to deliver uniform networking and security be made available in public clouds as well. This, of
controls across multi-cloud environments. This was course, allows customers to experience a consistent
extended in November 2019 as threat prevention look and feel even when they expand into public
was introduced as part of NSX, a feature that packs clouds, all using the VMware technology stack.
more of the previously mentioned “stages” into the
hardware abstraction layer, as depicted in Figure 2.

NSX-T WITH SECURITY

Host 1 Host 2

VM1 VM2 VM3 VM4 VM5

pg 10 pg 11 pg 10 pg 11
VLAN 10 VLAN 11 VLAN 10 VLAN 11 Virtual Switch

VMware Hypervisor Security VMware Hypervisor Security

Physical Network

Figure 2: NSX-T with Security

Now let’s look at the architectural challenges •Security controls run only on VMware hypervisors,
imposed when moving a richer set of security which requires the use of Amazon and Azure
controls into the hypervisor: proprietary hardware carveouts that are running
VMware technology stacks. Those stacks are
•Most of the same problems that monolithic
different from the cloud provider offerings.
traditional security VMs experience apply here as
Customers must pay for cloud providers and
well (scaling, HA, upgrades). Hypervisors are bound
VMware as well as incur a different feature set than
to certain compute, storage and memory limits
offered by clouds
available on the host and compete with workload
VMs for those provisions. Especially hard to predict •Traditionally, hypervisors were just focused on
are the capacity demands of threat prevention as hardware abstractions to keep the code minimal
they are a function of policy and the actual traffic and robust. Extending the footprint of a hypervisor
observed. This can lead to starving workloads or also extends the attack surface and carries the
security controls potential for stability issues as well as breaches
HYPERAUTOMATED AI-BASED SECURITY VS VMWARE NSX | 4

To sum it up one could argue that the NSX-T Figure 3 depicts how only a tiny presence per
solution resembles a physical network security host is required for the initial processing stage.
appliance running a hypervisor with security that We call this an FV for FlowVisor. A FlowVisor
leaves some space for other workloads to run. does to network flows what a hypervisor does to
hardware: it abstracts and unifies packet reception
Contrast the VMware solution with a ShieldX
and transmission independent of the underlying
ESP Hyperautomated AI-based security approach.
networking architecture. By making that FlowVisor
The art and science of looking deep into network
smart and intent-driven, it knows what traffic to
traffic flows is CPU and memory intense, whereas
process and how deep in a flow to go. It can send
just packet reception and transmission requires
that traffic on to subsequent stages that sit in the
lots of network IO. To devise the ideal solution, it
Elastic Backend. The backed can contain pooled
is important to realize that each individual traffic
sets of instances representing dynamically scaled
processing stage requires different scaling, and its
sets of stages. That pool enjoys the same luxury of
efficacy depends on varying hardware capabilities
modern virtualized or containerized infrastructure
like CPU, network IO or disk IO. Further, the
as the workloads they protect. This approach truly
amount of data exchanged decreases significantly
distributes all processing stages throughout the
the deeper the analysis steps go, thus making it
VMware data center, Amazon VPC, Azure VNET
possible to allow for pooling later stage security
infrastructures, etc., and does not impose the need
services across hosts. With that in mind we can
for a uniform hypervisor.
already guess what an ideal architecture looks like
Hyperautomated FlowVisor
that will profoundly overcome the previously listed
shortcomings.

HYPERAUTOMATED FLOWVISOR Host m-n

Elastic Backened
(N+1 HA)

Host 1 Host 2

VM1 VM2 VM3 FV VM4 VM5 FV

pg 10 pg 11 pg 10 pg 11
VLAN 10 VLAN 11
pg 12
VLAN 10 VLAN 11
pg 12 Virtual Switch

VMware Hypervisor VMware Hypervisor

Physical Network

Figure 1: NSX-V with Traditional Security Appliance


HYPERAUTOMATED AI-BASED SECURITY VS VMWARE NSX | 5

This architecture combines the FlowVisor with a •Native integration with Amazon EC2 and Microsoft
scalable Elastic Backend that is tightly coupled Azure not requiring proprietary hypervisors that are
with changes happening in the environment for incompatible with native cloud APIs
continuous adaptation. As such it outperforms
•Providing superior security to its own infrastructure
any other monolithic design as it delivers the
as it has two shells around each component: the
following advantages:
VM, and containers within VMs
•Ability to transparently insert, agent-less, into
•More effective security controls as they can share
the network traffic without disruption to the
state across instances. Other approaches would
applications and hypervisors but with automation
require state synchronization across hypervisors
and no user intervention
and 3rd party appliances, which is very hard to
•Auto-scaling out and in on an as-needed basis for accomplish if at all attempted
only the stage that needs help or is underutilized
To summarize, the ideal approach is to decouple
•N+1 High Availability vs. 2N in traditional models network security from the underlying hypervisor.
(provision only one extra for each stage) This delivers most reliability, effectiveness, flexibility
and agility to manage the solution over time.
•Most efficient resource consumption (no need for
Especially as public cloud providers expose APIs
having all stages on each host)
for network traffic monitoring, they natively built
•Automated and orchestrated policy computation the underpinnings of the FlowVisor and there is
and generation by learning groups and policies no driver to unify a technology stack across clouds
using ML for security. The proposed solution herein helps
customers additionally by capturing their intent
•Non-disruptive upgrades by supporting and configuring components automatically to
rolling upgrades achieve the intended security posture continuously
and optimally.
•Richest set of security controls easily extensible
without compromising hypervisor security, capacity
or safety

ShieldX Networks, Inc. +1 408.758.9400


2025 Gateway Place, Suite 400 info@shieldx.com
San Jose, CA 95110 USA www.shieldx.com

© 2018 ShieldX Networks, Inc. All rights reserved. For trademark, copyright, patent, November 2018
and other intellectual property and legal information, visit www.shieldx.com/legal

You might also like