Does Security Belong Near Endpoints?

Peter Welcher
Architect, Operations Technical Advisor

For a while now, I’ve been mulling over security architectures. I’ve started wondering if security really belongs on or near endpoints (users, servers, VMs). And are firewalls obsolete?

We need to make a distinction here. I rarely see edge firewalls as a problem. Internet speeds are just generally not that challenging for firewalls. Currently, anyway, although higher speeds (e.g. N x 10 Gbps) get costly.

This blog is more about datacenter core firewalls, and firewalling user segments — places where firewall throughput and cost are more likely to be problems. But it also relates to how we firewall an increasingly cloudy world going forward.

This blog is intended to ask questions, stimulate thought and discussion, not necessarily provide complete answers.

Why Firewalls?

The original problem was something along the lines that application developers don’t reliably document open ports and facilitate hardening servers, and server admins couldn’t be guaranteed to do the hardening, either not caring or for fear of breaking the application. Putting a firewall in front of the servers let network admins limit what source IPs and destination ports could be accessed. At least in principle.

Then firewalls “grew up” and became Next Generation: IPS / IDS, Application Layer Gateways (ALG), and other functions came along, first as dedicated boxes, more recently as Next Gen Firewalls.

As I’ve noted elsewhere, if you want access lists, switches do that cheaply and at wire speed.

Presumably firewalls are also there for stateful TCP enforcement, maybe some protocol correctness enforcement, and the IPS / IDS / ALG roles. Also, maybe to collect flow data, although some switches can do that as well.

I also see sites doing things like trying to run all user to user traffic through firewalls (segmentation and malware spread prevention). Or all WAN or VPN traffic.

When doing design / architecture, it helps to consider pros and cons.

For firewalls, the pros and cons that come to mind are:

  • Pro: Firewalls can act as great single points of security control and enforcement.
  • Pro: They give the security team something they solely control. This can be good, as in securing admin control over what traffic is allowed, but it can also be a form of people defining their job by the devices or tools they “own”.
  • Con: Firewalls do a lot of processor-intensive work, so higher-speed firewalls are expensive (and may be falling behind as 40 and 100 Gbps speeds become affordable).
  • Con: They don’t see all traffic (hence we have various flow tools getting into the security game). Cisco Tetration, for example.
  • Con: For the firewall to see, enforce and protect traffic, it has to go through the firewall.
    • Con: Network plumbing to use firewalls can be problematic (example: routing between two VRF segments at a site having to traverse the WAN to be centrally firewalled, due to trying to hold the firewall count down).
    • Con: Running traffic through a firewall increases the load, which can dramatically increase the cost of the firewall. And note you do need to size firewalls right. Put differently, there is a size range where you may can afford a certain firewall model and that firewall can do what you need. Above that, you will either have to dial back on what you’re asking the firewall to do, or spend more.


What are the alternatives? Well, that depends on what problem you’re trying to solve.

The big driver that comes to mind is price. Performance comes in when you’re talking firewalling in the middle of a datacenter using 40 Gbps links. 100, 400 Gbps???

Really, the only option is trading scale-up for scale-out. That may improve performance but may not improve price much (from having priced out virtual firewalls on VMware not that long ago).

Scale-out could mean an IDS that feeds data to a bank of compute engines. There may be other clever approaches. Clustering physical firewalls is more of an intelligent brute force scale-out.

That leaves us with some other approaches, or types of approach:

  • Cisco ACI micro-segmentation (is ACI “near” the endpoint?).
  • VMware NSX micro-segmentation (policies in VMware, can match on tags and VM names).
  • Server / VM agent-based solutions, e.g. Illumio.
  • Linux iptables based security (and automation of that across servers and VMs?).
  • Endpoint (user) solutions, e.g. (in part) Cisco AMP, other products.
  • Cloud gateway rules, to coin a name. Basically, ACLs for access between outside and virtual nets, and between virtual nets. Simplified virtual firewalls?
  • Virtual firewalls (as VM or Cloud Instance). These are basically like the physical firewalls, but likely lower throughput. That can be solved via scale-out depending on your single-flow throughput needs and your wallet.
  • Container-centric security, e.g. Tigera Calico, etc.

Some general comments:

  • I’m fine with ACI for physical endpoints. For heavily virtualized environments, maybe less so. ACI distributes enforcement to leaf switches, so it does scale-out.
  • VMware ACLs are naturally closer to VMs.
  • For cloud, I have the impression both ACI and NSX-T basically run traffic through a VM virtual L3 switch with controller-based forwarding rules. It does enforcement. Efficiency of that?
  • Agents: some people are allergic to them. Concerns include adding to the load on the server or VM, plus the agent breaking the application(s).

Here are some pros and cons thoughts (hard when there are so many choices):

  • Pro: The above are scale-out approaches.
  • Con: But how do I know that all my servers or user endpoints have the agent?
  • Dunno: Cost. Varies. I’m sure a sales person will be glad to take up your time, ahem, help you with that.
  • Concern: Central management. Without good central control, management, and reporting, agents or other distributed solutions can become quite labor-intensive.
  • Con: Yes, but (laws, corporate policy — choose one) say I must have a firewall.

Other Factors

Endpoint security might solve potential asymmetric path issues with Data Center Interconnect (DCI), LISP and stateful devices, especially firewalls. If the only stateful firewall is at the endpoint, then the path you take doesn’t matter, and asymmetry is not a problem. This might help with the WAN. Of course, how many sites are likely to connect to the Internet without a firewall? Stateful load balancers, especially those doing NAT — probably still a problem.

As noted above, making sure all your servers or user endpoints have the agent on them could be a concern. While it should be fairly easy for a product to report endpoints that do contain the agent, spotting agent-less endpoints does appear to be a harder problem. Sounds like a job for Cisco ISE!

This also brings us back to the start of the article: covering for devices lacking security. Some might argue for a firewall as a form of “belt and suspenders” insurance.

In my reading about containers, I’ve come across the “sidecar model”. The short version is that instead of agents, the automation tool / service mesh framework puts a “sidecar” inline for traffic going to your containers. Think “agent in a separate container” if you want. Or “mini-firewall” in a container?

The attraction of that approach is that the service automation tool can ensure uniform or controlled deployment of sidecars, i.e. no gaps in coverage. And less chance of breaking the code in the container, since the sidecar only filters traffic — no kernel or other hooks in the application container.

Yes, there is cost to doing sidecars — more compute, possibly some latency. If there’s a chain of sidecars (Is that a train?), figuring out which one ate your packets might be interesting, not fun.

Some Other Firewalling Approaches

Google shifted to using device and identity-based or context-based security, called BeyondCorp. That’s another approach, and controls who can get to what. It is available in GCP as context-based access via Identity Based Proxy.

Looking ahead (or now), there’s also the elephant in the room: IOT. Right now, using segmentation and IOT gateways is one approach. IOT gateways are there perhaps for a reason similar to what we started with for firewalls: managers have no idea what the vulnerabilities might be, so they want to intercept everything to control what goes to or from the IOT devices.

Putting agents on IOT devices isn’t likely to happen. Sidecars only work for containers, although the approach might be viable for Cloud providers to supply for VM instances.

Random thought: What happens (in 20 years?) when APIs drive all computer-to-computer traffic to be HTTPS-based? Including DNS?

As I was writing this, I ran across an Apstra blog on micro-segmentation. It suggests separating intent (policy) from where it is enforced, with IOT as a use case. That’s a valid point, near the endpoint may not fit all situations. Or will IOT change over time, to where power and other budgets allow room for agents?

Extra Credit Question

As we move from VM instances in the Cloud to Containers then serverless, or use different Cloud tools for different business reasons, what does security for serverless look like?


To NetCraftsmen’s Samuel Bickham, for reviewing this. Any errors are mine!


Comments are welcome, both in agreement or constructive disagreement about the above. I enjoy hearing from readers and carrying on deeper discussion via comments. Thanks in advance!


Hashtags: #CiscoChampion #TechFieldDay #TheNetCraftsmenWay #Firewall #DataCenter #CyberSecurity

Twitter: @pjwelcher

Disclosure Statement
Cisco Certified 20 Years

NetCraftsmen Services

Did you know that NetCraftsmen does network /datacenter / security / collaboration design / design review? Or that we have deep UC&C experts on staff, including @ucguerilla? For more information, contact us at

Leave a Reply