Slashdot Mirror
Amazon Plans To Challenge Cisco in Networking Market With Much Cheaper Switches, Report Says (theinformation.com)
Amazon Web Services already dominates the market for cloud services. Now, reports The Information, it is eyeing a part of the cloud business it doesn't already control: the $14 billion global market for data center switches [Editor's note: the link may be paywalled; alternative source]. From the report: AWS is considering selling its own networking switches for business customers -- hardware devices that move traffic around networks, according to a person with direct knowledge of the cloud unit's plans and another person who has been briefed on the project. The plan could plunge Amazon more deeply into the lucrative enterprise computing market, posing a direct challenge to incumbents in the business like Cisco, along with Arista Networks and Juniper Networks.
As it does in many other categories, Amazon plans to use price to undercut rivals. The company could price its white-box switches between 70% and 80% less than comparable switches from Cisco, one of the people with knowledge of the program estimated.
As it does in many other categories, Amazon plans to use price to undercut rivals. The company could price its white-box switches between 70% and 80% less than comparable switches from Cisco, one of the people with knowledge of the program estimated.
Cisco is just power of a name soon a new name will step up and challenge
This is why Cisco purchased (2003), absorbed, destroyed, and released (2013) Linksys - their higher end devices were able to replace a growing percentage of the switches and routers being marketed towards smaller businesses. M&A is a very successful way to kill a competitor in the US, GOV rarely cares and is for sale, and the investors rarely care after they cash out. But Cisco can't afford Amazon. High end switch market has been a mess, software configured networking is eating it alive, and its amazing what you can do with a simple Docker network. Be nice to see someone with a budget release some cheaper hardware where we still need actual hardware.
Cisco always has to prices: The list-price and the retail price to customers who are "in the know". The latter is usually 60-70% below list-price.
Which one is Amazon going to undercut ? If it is the first... Meh... Not so interesting.
If it is the second... Then things get interresting. They will even be undercutting HPE/Aruba then.
We are able to get switches and routers for cheap for a while. Many have the same features that Cisco offers.
The reason most companies stick with Cisco, is because they are able to find Certified Staff to work on their products.
If a company tried to upgrade to Amazon Fire Sale Switches, then you need to find staff willing to maintain them and do it properly with best practices in mind, may be difficult. You can probably get Cisco Certified staff to work on them, however if there are any differences there may be an issue.
If something is so important that you feel the need to post it on the internet... It probably isn't that important.
Privacy concerns will be massive!
True, but they have a huge competitive advantage with linking that name and their very popular cloud service. At some point, the regulators will start taking a more active interest in Amazon, but Iâ(TM)m sure Amazon already has a significant number of lawyers working on this very issue.
If you want news from today, you have to come back tomorrow.
Prime members have their packets delivered in 2 nanoseconds or less.
From Amazon's perspective this makes sense, provide priority bandwidth for Alexa and Amazon Prime as well as providing a way of monitoring customers' internet habits. Hopefully, they will be providing a high level of security so the information they're accessing/collecting doesn't become available to third parties.
When there are *lots* of low-cost switches that I don't have to worry about Amazon's potential for taking over my home, why would I want to buy from them?
Mimetics Inc. Twitter
I.e to get into market, start with solid cheap stuff (where the requirements are low). Then try to fight the big players.
My estimated outcome: either they do not survive one year or they become as expensive as Cisco if they win.
Is it still going to be cheaper than Cisco when you pay to not get ads delivered to everything connected to your network switch?
Is this going to be like their phones and their tablets and their e-readers where you have to pay more not to get ads?
"That's the way to do it" - Punch
100 years from now Amazon will be the new Umbrella or Cyberdyne System Corporation.
They're only 5 years away from being Veridian Dynamics.
"That's the way to do it" - Punch
Software defined networking is great when dealing with networks at a high enough level. People have been making routers from commodity hardware for a very long time. Obviously people have produced special purpose hardware for routing as this means they can optimize the hardware for the task and can do so cheaper than someone grabbing a PC, filling it with interface cards, and loading some software onto it.
Switching is different than routing, it's done on a different level. The hardware needed is more complex, and therefore more expensive, than what is found in commodity computers. Go and try to find a software defined switch. I tried, and they don't exist. The closest you will find is a switch defined as a virtual machine. Load up something like VMWare ESXi and you'll find a way to create a software switch, but it can only switch packets among the virtual machines on that system.
People have made limited software switches with server style Ethernet cards (which grant greater access to the packet content than a desktop Ethernet controller) and the right kind of software but they are expensive and slow. They are really only useful for things like testing, training, or demonstrations.
This is a big deal because this means Amazon is getting in the hardware business in a way that is quite rare. Amazon is a large enough company that they may actually be able to follow through.
I am armed because I am free. I am free because I am armed.
Well, I was assuming the parent was joking, that it could be software all the way down, which isn't obviously possible.
In terms of software defined switches, generally speaking they consider any switch that can be ONIE to be 'SDN-friendly', and some others.Sure, there are switching chips doing the actual moving of the data (there pretty much has to be), but their primitive capabilities are exposed to software for more in depth wrangling.
In practice though the complexity of SDN switching is well beyond the point of diminishing returns for almost everywhere to bother with.
XML is like violence. If it doesn't solve the problem, use more.
I heard the same thing about John Deere growing up. That John Deere was just green paint and a lot of snobbery.
Here's what I learned, farmers and ranchers are businessmen. They need to get work done like every other business. Downtime costs money. John Deere tractors still break down, get stuck in the mud, wear out, etc. It's that the competition do this more often. There's still some snobbery and such in there, John Deere tractors can have leather seats and built in refrigerator. They spend the money on the "green paint" because it gives them more return on their investment.
Is Cisco just a name? Maybe that's true now but they can only get to be "a name" by proving to be better over time. No one Is GMC just a name? Is Apple? Businessmen buy this stuff because it makes them money. If Cisco stops making people money, or rather they can make more money with someone else, then Cisco will disappear. Same goes for Apple, John Deere, and GMC.
I am armed because I am free. I am free because I am armed.
"Hardware? I though everything was headed towards SDN (software defined networking)?
https://en.wikipedia.org/wiki/Software-defined_networking"
(Your article says SDN started becoming a topic in 2011.)
1)You would think AWS would know a bit about SDN, since they basically invented it, and made it available as part of a service to customers in 2009, except they called it VPC:
https://en.m.wikipedia.org/wiki/Amazon_Virtual_Private_Cloud
2)You probably still need networking-specific hardware; you want a higher ratio of network ports to CPUs in a typical switch than a typical PC or server. Amazon already makes custom network hardware for servers and switches using 25GbE, apparently including 100GbE using QSFP (4Ã--25GbE) (where AFAIK Cisco only does up to 4x10GbE between switches and their server hardware and 100GbE is only available between reall really expensive core switches - N7K - and really really expensive routers - ASR9K or NCS6K).
See e.g.
https://www.theregister.co.uk/2016/11/30/aws_hardware/
It won't take a whole lot to undercut Cisco since they have always had ridiculous pricing.
Even companies with damn near infinite amounts of cash finally started looking at other vendors because of ludicrous price levels.
However !
That said, I have decommissioned Cisco routers and switches that have been running ( without a reboot ) for twenty plus YEARS without a hiccup.
I doubt you're going to find that sort of reliability in anything offered at rock bottom prices.
So, while expensive as hell, I can't complain about the operational track record.
Hmm, I'm using 10Gbps software switch as real 10Gbps switch is still too expensive...
Just group your interfaces into a bridge -- after all the core function of a switch is a bridge. No software required; it's build into the kernel already.
# man bridge
Because fuck you, that's why.
They're so locked into that mentality that 10 Gbps needs to cost $$$$ that instead of pushing forward they're pushing backward by shitting out multiple half-assed "solutions" for teaming two 1 Gbps links on consumer / small business gear, as well as teasing eventual support for 2.5 Gbps and 5 Gbps standards.
They blame cat 5e cabling, but you can run 10 Gbps over cat 5e in "short" runs (probably up to 40 meters).
So what are Amazon's plans to address support of both software and hardware? How will they fix security issues and other bugs? Will they provide the needed assurance this won't be abandoned in another couple years for the next new shiny thing? Or do they just plan to dump their stuff on the market and hope for the best and tell the end customer they need to support it?
So what are Cisco's plans to address support of both software and hardware? Shit out untested, buggy software with huge security holes on reliable, but extremely overpriced hardware? Etc.
Cisco isn't exactly highly regarded anymore. It's the entrenched standard people are afraid to move away from. The devil you know.
Switching is in fact FAR simpler then routing. Layer 3 switches are a hybrid router with switching logic. Pure switching simply looks at feild X in a packet and switches it to the correct egress port with a single table lookup. That is trivial. Routing has to look at the entire table and match based on a list of rules. Layer 3 switches let you bring the joys of policy routing to switching. There is real need for 100G switches in a affordable price point for DC's. Along with that they want layer3 switching/routing.
Well, I was assuming the parent was joking, that it could be software all the way down, which isn't obviously possible.
Yes, I realized that was quite probable after I submitted my post.
In practice though the complexity of SDN switching is well beyond the point of diminishing returns for almost everywhere to bother with.
Agreed, I imagine there is a market for software defined switching but it is quite small because the costs outweigh the benefits for most cases. I can also imagine much of that market exists in places where much of the network is virtual, like the VM clusters I mentioned in my previous post. It may be possible that software defined switches could gain more of the market. I'm thinking that not only would cost be a consideration but also security. I don't know much about how software defined switching would work but I'm quite certain the more general purpose the hardware the less secure it will be.
I am armed because I am free. I am free because I am armed.
If Amazon uses their own switch to power their own super huge datacenter... that's a solid argument that those switches actually works!
At layer 2, the promise of value is more granular control over packet forwarding than designating vlans.
The switch chips under the covers have a great deal of impossibly complicated capabilities that traditional switch config software abstracts away to basically vlan and not much else. Traditionally there is also sometimes helpful filtering (e.g. 'do not forward ethernet frame if it's dhcp response'), though that is a bit rare and generally hard to configure. There exists a contingent of folks who want to go deeper and imagine higher performance topology (e.g. a fat tree, torus, dragonfly, basically the sorts of topologies you see in infiniband and omnipath) that spanning tree would spit all over, and MST or similar would be too coarse. TRILL was the 'non-SDN' answer proposed to provide other topologies on ethernet, but that didn't pan out.
Problem is that in practice, it's trying to reinvent the infiniband sort of strategy (openflow controller is like an infiniband subnet manager) and this is very difficult to pull off, and generally superfluous for most people and the rest could... just get infiniband where the solution is pretty mature....
XML is like violence. If it doesn't solve the problem, use more.
Software defined networking is only useful for service providers to secure access. When you have multiple customers hosted on the same VMWare servers you need to make a network that is scale-able and secure. You do this by creating a separate subnet and a private vlan. If they have multiple machines across several vmware hosts then you create a community pvlan and away you go. Everyone can share the same primary private vlan which allows for easy subnetting but the switch won't allow them to cross secondary vlans. This is all defined in-software, at the switch level it is just a single tagged vlan primary vlan, and then tagged secondary vlans for each customer.
SDN is just like using the term cloud, it's just another term for something people have been doing for over a decade. If you're a business with only a few VLANs and not much infrastructure churn then you have no need for SDN. If you have a lot of churn then it can make sense as it is easy to scale out and remain PCI, hippa, or ferpa compliant.
In my environment the first customer took me a week to provision properly with SDN, the second customer took less than 5 minutes. I'm limited only by my ability to spin up a new vm from my gold image and then assigning it to the new network. Everything else takes about 1 minutes start to finish.
Yes this can be done with a regular computer but it will not perform at near the same level as a real managed switch that is using ASICs to do all the work which are purpose built to do exactly that. They will outperform a CPU doing the work every-time. You can throw a ton of CPU at a PFSense box and achieve good performance but then you might as well bought a real firewall which will be easier to manage and perform even better.
I say that as someone that threw together two old servers to make a PFSense HA cluster until we could afford actual firewalls which more than doubled our performance overnight. It's easy to look at a NIC and think that its 10gig so put two together and you'll get 10gig switching throughput. You'll be lucky to get 5 out of it.
Well, at least until their recent fascination with DRM in their equipment.
The NSA was already caught hacking Cisco's routers before foreign customers received them. I wonder how secure Amazon's are? Do they subcontract the manufacture to China? Does the Chinese government get a back door out of the deal?
Yes, software switches do exist (aka "bridge"), but, as you mention, they're slow as crap because software (general purpose CPU) has to move frames from interface to interface.
Amazon isn't "getting into the hardware biz". They're just going to (sub)contract that shit to any number of "white box" switch makers already gluing common Broadcom (etc.) switch SoCs to boards. The OS on those boxes will most likely just be a customized / rebadged existing network OS.
Says the guy that has probably never needed to call Cisco TAC lvl2 about a zero day bug found in a $10 mil+ DC.
There's lots of competitors out there and the telco I work for has all of them in small amounts in both DC's but none of them comes close to Cisco TAC until that changes we won't move over anything major over to the wannabe's except edge devices (Aruba switch/controller/ap combo's).
Juniper comes the closest but there layer7 support (next gen firewalls like Palo's) isn't there yet.
Cisco maintained their position largely due to their upstream training and certification. They haven't been any better then the competition and they are worse in many ways.
Cheap storage VM.
I donâ€(TM)t see this being a problem.
:
Ciscoâ€(TM)s data center switches (something which has fed me and my family for 6 years) are not adapting to modern networks. Cisco is so heâ€(TM)ll bent on ACI and even EVPN that they are not making their systems cloud friendly.
See, ACI is â€oeSoftware Defined†in purely the loosest sense of the word. It is very poorly suited for use with containers and FaaS as those systems leave most of the networking to systems like Kubernetes and the ACI topology isnâ€(TM)t well suited for those topologies.
EVPN is nifty if you need a lot of layer-2 broadcast domains that terminate at physical ports. But containers and FaaS terminate inside of Docker for example.
Legacy data centers thrive on high performance low latency links. This is because of two main features. The first is storage and the second is virtual machine migration (vMotion for example). Systems controlled as containers have substantially lower bandwidth requirements since storage is far smaller (30-100MB containers vs. 40GB or larger virtual disks) and because we distribute the containers predictively, we can do it far slower. We donâ€(TM)t migrate virtual machines either.
In a well designed container platform most database actions are performed with Map/Reduce technologies. This means the only traffic on the data center network is query and result. We donâ€(TM)t attempt to read terabyte or larger files from storage systems. We instead transmit a query to the nodes containing segments of the data and collect the results of the query and reduce duplicate responses. This does benefit from low latency, but high speed (10GBe, 40GBe, or greater) has no real performance benefit.
Cisco QoS is as always based on hardware, this limits the QoS mechanisms to effectively a small number of queues. Prioritization is limited as there are effectively 6 usable classes of traffic. While drop probability in DSCP can be helpful, itâ€(TM)s very difficult to implement meaningfully when the network canâ€(TM)t understand the actual type of data involved.
A proper data center switch would be fully programmable on a stream by stream basis. Like the back end of ACI or FabricPath, it would break from traditional Ethernet forwarding and instead use traffic specific tunnels with real understanding of QoS needs. This canâ€(TM)t be done with Cisco hardware.
An optimal data center switch topology would have the following
1) High performance later-3 switches for legacy virtual machine support. Simple IPv4 routing with large buffers and marking for low-latency lossless would suffice for almost all data center needs. NSX and Hyper-V will handle the rest.
2) 1Gb or multi-gig (latency not performance) with enormous buffers for all modern container and FaaS traffic. They should be able to have extremely high performance REST APIs to insert and purge streams into/from the topology with QoS rules. They should be entirely layer-3 based and should allow Docker/Kubernetes or others to program MAC address tables and should block all layer-2 traffic which hasnâ€(TM)t been programmed into the forwarding table by a controller.
Now that being said, pure layer-3 switching with NAT support would be far better. Layer-2 is dead. All of that can be programmed from the control plane and skip learning. This isnâ€(TM)t 1990 when every machine had a random MAC address which had to be learned accidentally. Instead, data centers and clouds (container farms and FaaS) know all the MAC addresses of all the interfaces properly. In fact, the controllers already have all the IP to MAC mappings known internally. As such, a data center switch would allow these to be programmed instead of learned or snooped.
So the way I see it, Amazon is on the
I'd love to see this effort drive down the costs of 10gb+ network speeds and drive them into the consumer market.