"Cloud Networking" defined in Arista Networks new CEO Jayshree Ullal's blog

[Pranta Das]

http://www.aristanetworks.com/en/JU_Cloud_Networking

[Tarry Singh]

With Silicon Valley taking a much larger hit at funding and staff lay-off, is it a good idea to do the start-up? We did follow that Andy left Sun to do this. Although from another perspective, this is the moment to lay seeds for the new order as well. Nice buzzwords and shiny balls (http://www.aristanetworks.com/en/Solutions), what else are they doing besides 10GbE and PHY?

I think I need to dig in deep and see what they do.

Tarry

On Thu, Oct 23, 2008 at 5:52 PM, Pranta Das <pran...@yahoo.com> wrote:

http://www.aristanetworks.com/en/JU_Cloud_Networking

--
Kind Regards,

Tarry Singh
______________________________________________________________
Founder, Avastu: Research-Analysis-Ideation
"Do something with your ideas!"
http://www.avastu.com
Business Cell: +31630617633
Private Cell: +31629159400
LinkedIn: http://www.linkedin.com/in/tarrysingh
Blogs: http://tarrysingh.blogspot.com

[Sundeep Singatwaria]

Tarry -
Dense 10 Gig switches would work very well for cloud networks. The QoS model in the enterprise world never took off. Network administrators are used to throwing more bandwidth at the problem of deteriorating Quality of Service for user's traffic. So packing more 10 gig ports in a 1 RU box makes all the sense - if you follow the problem from traditional networking folks point of view.
 
People who do application acceleration / optimization networking would want to use Content Aware Switches. These switches can classify user traffic using Deep Packet Inspection (DPI) engine. They can apply Auto QoS based on what the user is trying to do. Auto QoS can be derived from pre-defined profiles for an organization and its people being serviced by the cloud. For a user connecting to the cloud - he /she may get different QoE (Quality of Experience) based on what their role in the organization is and/or what application they are trying to access.
 
Dense 10 Gig doesn't require cloud network administrator to do elaborate configuration. However Content Aware Switches will require elaborate configuration - Company / User / Role to QoS profiles binding.... The whole SLA is more difficult to manage with Content Aware Switches.
 
If AAA + A is done within network infrastructure then it is easy to deploy Content Aware Switches since QoS can be deployed alongside AAA + A configuration at no extra cost. However if AAA + A is deployed within application then dense 10 Gig switches will make all the sense.
 
Hope you follow all this cool stuff !

---

Date: Fri, 24 Oct 2008 13:40:18 +0200
From: tarry...@gmail.com
To: cloud-c...@googlegroups.com
Subject: Re: "Cloud Networking" defined in Arista Networks new CEO Jayshree Ullal's blog

[Chris Marino]

It's been a while since I really looked into these things, but I seem to recall one of the things holding back 10GE was the actually filling it up.  TCP itself becomes a bottleneck with any reasonable sized transfer. Lots of work was done with TCP offload engines (TOEs), fast restart, buffer size, etc. but I don't think they ever really solved the problem.  I'd have to go look at the current disk performance stats, but I doubt that raw disks could even fill it up....

 

I'm sure high density 10GE switches will be common place before too long, but don't really see them as a game changer.  Can't see my EC2 images coming anywhere near maxing out it's interface anytime soon.

 

Also, Not sure I buy the comment (in NYT article from Andy B.) that the 'iPhone runs better software than a typcal switch'.  Might be true of the content aware switches, but it's a stretch to consider them 'typical' by any measure.

 

One final point: I totally agree w/Sundeep in that 'content aware switched require elaborate configuration'.  So which cloud providers are going to let me configure that? 

 

Probably the same ones that let me config their routers...

 

CM

[Chris Sears]

I've heard anecdotal reports that many times the network can be saturated on an EC2 instance, especially if you happen to land on a box with other heavy network users. Same with disk I/O... o the degree that some have found EBS is actually faster than local disk.

I think the promise of 10GE is a single, unified fabric that I could use for storage (FCoE or iSCSI) in addition to normal production TCP/IP uses. Cisco is moving in that direction with their Nexus line. Outside of cloud computing environments, I think anyone not stuck with an extensive FC fabric/infrastructure will be seriously considering 10GE as an alternative with some room for future growth.

And getting back to the EC2 example, cloud providers need a network infrastructure that can keep up with the advancements happening inside the servers... with hexa- and octal-core CPUs just around the corner and RAM constantly getting faster and cheaper, providers are looking at hosting many more VM instances per box. And with EBS / iSCSI / FCoE in the picture, the network traffic starts to add up pretty fast. We're going to need a very fat pipe going into each physical server to continue scaling for the next 5 years. 10GE looks like a good fit.

 - Chris

[Laurent Therond]

I heard TCP is no good, but I also heard 99% of everything out there runs
on top of it.

</sarcasm>

Hardware vendors and software vendors will continue to compete.

> Date: Fri, 24 Oct 2008 13:40:18 +0200
> From: tarry...@gmail.com
> To: cloud-c...@googlegroups.com
> Subject: Re: "Cloud Networking" defined in Arista Networks new CEO
> Jayshree Ullal's blog
>
> With Silicon Valley taking a much larger hit at funding and staff
> lay-off, is it a good idea to do the start-up? We did follow that Andy
> left Sun to do this. Although from another perspective, this is the
> moment to lay seeds for the new order as well. Nice buzzwords and shiny
> balls (http://www.aristanetworks.com/en/Solutions), what else are they
> doing besides 10GbE and PHY?
>
> I think I need to dig in deep and see what they do.
>
> Tarry
>
>
>
> On Thu, Oct 23, 2008 at 5:52 PM, Pranta Das <pran...@yahoo.com> wrote:
>
>
> http://www.aristanetworks.com/en/JU_Cloud_Networking
>
>
>
>
>
>
>
>
>
>
> --
> Kind Regards,
>
> Tarry Singh
> ______________________________________________________________
> Founder, Avastu: Research-Analysis-Ideation
> "Do something with your ideas!"

> http://www.avastu.com <http://www.avastu.com/>

> Business Cell: +31630617633
> Private Cell: +31629159400
> LinkedIn: http://www.linkedin.com/in/tarrysingh

> Blogs: http://tarrysingh.blogspot.com <http://tarrysingh.blogspot.com/>
>
>
>
>
>
>
>
> >
>

[Chris Marino]

-----Original Message-----
From: cloud-c...@googlegroups.com [mailto:cloud-c...@googlegroups.com] On Behalf Of Chris Sears
Sent: Friday, October 24, 2008 1:29 PM
To: cloud-c...@googlegroups.com
Subject: Re: "Cloud Networking" defined in Arista Networks new CEO Jayshree Ullal's blog

I've heard anecdotal reports that many times the network can be saturated on an EC2 instance, especially if you happen to land on a box with other heavy network users. Same with disk I/O... o the degree that some have found EBS is actually faster than local disk. 

 

Bottlenecks are slippery little devils....what you say may be true, but I've yet to see any data that tells me that it's the interconnect, and not some other chokepoint (i.e. Hypervisor, TCP,  SCSI, PCI, FSB, DDR, whatever...)

I think the promise of 10GE is a single, unified fabric that I could use for storage (FCoE or iSCSI) in addition to normal production TCP/IP uses. Cisco is moving in that direction with their Nexus line. Outside of cloud computing environments, I think anyone not stuck with an extensive FC fabric/infrastructure will be seriously considering 10GE as an alternative with some room for future growth. 

 

100% agree. I believe that iSCSI was the primary motivation behind a lot of the TOE work a few years back.... 

And getting back to the EC2 example, cloud providers need a network infrastructure that can keep up with the advancements happening inside the servers... with hexa- and octal-core CPUs just around the corner and RAM constantly getting faster and cheaper, providers are looking at hosting many more VM instances per box. And with EBS / iSCSI / FCoE in the picture, the network traffic starts to add up pretty fast. We're going to need a very fat pipe going into each physical server to continue scaling for the next 5 years. 10GE looks like a good fit. 

 

I/O has not kept pace w/CPU/Disk/network capacity and remains several order of magnitudes behind.  Fast multi-cores make going off-chip really expensive (i.e. slow). Spanning servers is even worse and lots of large workloads trade off I/O for compute whenever possible.

 

Of course 10GE is inevitable, but isn't not a game changer. I've got 100ME all through my house, but YouTube is about the same speed as my iPhone.....

 

I think this new switch is trying to compete on features, not performance. Which makes sense.  Not at all clear how those features will be available to users up the stack.

 

CM

 

 - Chris

[Krishna Sankar (ksankar)]

In one sense you are right - if we need to do application level complex configurations on routers, then the point is lost in many ways - first all the boxes in the path need to change and second the network admins would get very mad; very very mad ;o)

But the question to ask is - do we need to do complex configuration for "content aware switches" Actually no - it should come from intelligent protocols. The context should be inferred and declarative rules followed.

The app QoS is not that relevant for clouds imho; they are more useful for "network-y" domains like video and voice. Also explorations in the line of "what the 'user is trying to do'", "their role in the enterprise" et al are not really applicable in a cloud environment as there would 1000s of users in a cloud and so aggregate metrics, SLAs, mobility and deterministic latency are more important. The user experience is a second order result of the master SLA; moreover these come into play only during contention. Bandwidth in a cloud env should be abundant (while finite, of course) - most probably the cloud networks would look like core networks - with speeds and feeds. So as Sundeep says, dense 10Gps switches make sense.

Finally, the network does not have the full context - the layers at the VMs and beyond do; so either we need a control point above the VMs or we would need to capture the context via intelligent protocols that the network can understand and process, at which point a very fast network that also can parse the "cloud context" makes sense.

Of course there is more to the story ... and needless to say Cloud Networking is an interesting field

Cheers
<k/>

P.S: BTW, unfortunately all the above are my own (and only my own) beliefs and comments; and not necessarily my employer's !

From: cloud-c...@googlegroups.com [mailto:cloud-c...@googlegroups.com] On Behalf Of Chris Marino
Sent: Friday, October 24, 2008 12:27 PM
To: cloud-c...@googlegroups.com
Subject: RE: "Cloud Networking" defined in Arista Networks new CEO Jayshree Ullal's blog

It's been a while since I really looked into these things, but I seem to recall one of the things holding back 10GE was the actually filling it up.  TCP itself becomes a bottleneck with any reasonable sized transfer. Lots of work was done with TCP offload engines (TOEs), fast restart, buffer size, etc. but I don't think they ever really solved the problem.  I'd have to go look at the current disk performance stats, but I doubt that raw disks could even fill it up....
 
I'm sure high density 10GE switches will be common place before too long, but don't really see them as a game changer.  Can't see my EC2 images coming anywhere near maxing out it's interface anytime soon.
 
Also, Not sure I buy the comment (in NYT article from Andy B.) that the 'iPhone runs better software than a typcal switch'.  Might be true of the content aware switches, but it's a stretch to consider them 'typical' by any measure.
 
One final point: I totally agree w/Sundeep in that 'content aware switched require elaborate configuration'.  So which cloud providers are going to let me configure that? 
 
Probably the same ones that let me config their routers...
 
CM
 
 
-----Original Message-----
From: cloud-c...@googlegroups.com [mailto:cloud-c...@googlegroups.com] On Behalf Of Sundeep Singatwaria
Sent: Friday, October 24, 2008 11:28 AM
To: cloud-c...@googlegroups.com
Subject: RE: "Cloud Networking" defined in Arista Networks new CEO Jayshree Ullal's blog
Tarry -
Dense 10 Gig switches would work very well for cloud networks. The QoS model in the enterprise world never took off. Network administrators are used to throwing more bandwidth at the problem of deteriorating Quality of Service for user's traffic. So packing more 10 gig ports in a 1 RU box makes all the sense - if you follow the problem from traditional networking folks point of view.
 
People who do application acceleration / optimization networking would want to use Content Aware Switches. These switches can classify user traffic using Deep Packet Inspection (DPI) engine. They can apply Auto QoS based on what the user is trying to do. Auto QoS can be derived from pre-defined profiles for an organization and its people being serviced by the cloud. For a user connecting to the cloud - he /she may get different QoE (Quality of Experience) based on what their role in the organization is and/or what application they are trying to access.
 
Dense 10 Gig doesn't require cloud network administrator to do elaborate configuration. However Content Aware Switches will require elaborate configuration - Company / User / Role to QoS profiles binding.... The whole SLA is more difficult to manage with Content Aware Switches.
 
If AAA + A is done within network infrastructure then it is easy to deploy Content Aware Switches since QoS can be deployed alongside AAA + A configuration at no extra cost. However if AAA + A is deployed within application then dense 10 Gig switches will make all the sense.
 
Hope you follow all this cool stuff !

________________________________________

Date: Fri, 24 Oct 2008 13:40:18 +0200
From: tarry...@gmail.com
To: cloud-c...@googlegroups.com
Subject: Re: "Cloud Networking" defined in Arista Networks new CEO Jayshree Ullal's blog

With Silicon Valley taking a much larger hit at funding and staff lay-off, is it a good idea to do the start-up? We did follow that Andy left Sun to do this. Although from another perspective, this is the moment to lay seeds for the new order as well. Nice buzzwords and shiny balls (http://www.aristanetworks.com/en/Solutions), what else are they doing besides 10GbE and PHY?

I think I need to dig in deep and see what they do.

Tarry

On Thu, Oct 23, 2008 at 5:52 PM, Pranta Das <pran...@yahoo.com> wrote:
http://www.aristanetworks.com/en/JU_Cloud_Networking






--
Kind Regards,

Tarry Singh
______________________________________________________________
Founder, Avastu: Research-Analysis-Ideation
"Do something with your ideas!"
http://www.avastu.com
Business Cell: +31630617633
Private Cell: +31629159400
LinkedIn: http://www.linkedin.com/in/tarrysingh
Blogs: http://tarrysingh.blogspot.com

</HTML<BR

[Greg Pfister]

On Oct 24, 3:40 pm, "Laurent Therond" <cloudcomputi...@mailsieve.com>
wrote:

> I heard TCP is no good, but I also heard 99% of everything out
> there runs on top of it.
>
> </sarcasm>
>
> Hardware vendors and software vendors will continue to compete.

This is true, but misses a major point:

Existing lossy flow-control on Ethernet is not very good for disk
connections. Dropping frames and packets makes for pretty bad
performance.

There are two proposals to fix this, providing lossless flow control
on Ethernet.

One is from Cisco, who is doing its usual thing of putting something
out there and then working to have it accepted since it's "already" a
de facto standard (as in N Wifi). That's Data Center Ethernet (tm
Cisco).

Others are pushing the the standards groups to actual community
development of a standard, known as Converged Enhanced Ethernet, which
like DCE provides lossless flow control and multiple virtual lanes,
and, since they won't have it ahead of time, wouldn't give Cisco an
instant market advantage.

I don't know who's going to win or when. Whoever wins, the result is
Ethernet only in name and low-level physical interconnect. Everything
from there up in standard Ethernet assumes it can drop packets on the
floor when the going gets tough (congested). Lots of stuff being taken
from InfiniBand and ATM and reused here. New switches, etc., will be
needed to support it.

As someone said about Fortran: I don't know what the converged
interconnect will be, but it will be called Ethernet.

...eventually. This will happen; it's just a matter of time. But it's
not there yet.

Greg Pfister
http://perilsofparallel.blogspot.com/

> > From: tarry.si...@gmail.com

> > To: cloud-c...@googlegroups.com
> > Subject: Re: "Cloud Networking" defined in Arista Networks new CEO
> > Jayshree Ullal's blog
>
> > With Silicon Valley taking a much larger hit at funding and staff
> > lay-off, is it a good idea to do the start-up? We did follow that Andy
> > left Sun to do this. Although from another perspective, this is the
> > moment to lay seeds for the new order as well. Nice buzzwords and shiny
> > balls (http://www.aristanetworks.com/en/Solutions), what else are they
> > doing besides 10GbE and PHY?
>
> > I think I need to dig in deep and see what they do.
>
> > Tarry
>

[Rich Wellner]

Chris Marino wrote:

It's been a while since I really looked into these things, but I seem to recall one of the things holding back 10GE was the actually filling it up.  TCP itself becomes a bottleneck with any reasonable sized transfer. Lots of work was done with TCP offload engines (TOEs), fast restart, buffer size, etc. but I don't think they ever really solved the problem.  I'd have to go look at the current disk performance stats, but I doubt that raw disks could even fill it up....

Yup.  To a very rough approximation, one machine can feed 1Gbps.

UDT helps in general.

GridFTP helps for file transfers.

But, yes, it's a problem.

[reza moini]

Chelsio (www.chelsio.com) offers the best TOE solution.  Take a look.


 

---

Date: Mon, 27 Oct 2008 11:26:10 -0500
From: goo...@objenv.com

To: cloud-c...@googlegroups.com
Subject: Re: "Cloud Networking" defined in Arista Networks new CEO Jayshree Ullal's blog

[Sundeep Singatwaria]

 
I don't think cloud providers will allow individuals accessing the cloud to configure their boxes. That would be adminstrative nightmare for providers. What you get from a cloud provider would be based on your SLA with the provider.

 

---

</HTML<BR

[Jim Peters]

Some of you might find this a bit simplistic, but IMO if a cloud provider were to allow customers to configure their boxes, then it would be a grid, not a cloud. That's the whole point.

--
Jim Peters
+415-608-0851

[shreyas shah]

"EOS can distribute software and configuration information across the entire fabric, enabling seamless consistency across the infrastructure. This in turn enables to provision clouds, add new servers or services, and update images and patches with simplicity and ease, and to automate every step of the provisioning process"

 

Arista is talking about state less networking .....Any service (application) can be provisioned in the clouds almost instantaneously and send the configurations in network switches in the data center cloud.....

[Sundeep Singatwaria]

 
Krishna - Content aware switches would make sense for cloud providers who do not have deep pockets for "adundant" bandwidth. If you deploy content aware switches you only configure SLA on the box where people connect to - not on all boxes inside of your cloud. All the QoS and network security inference for remaining switches in the cloud will come from the content aware switch.  It is similar to the firewall world where people connect to the firewall to access company's resources. Not all switches and routers behind the firewall have firewall policies.
 
The app QoS in the cloud again would be a must have. Not only for apps like Voice and Video but also for everything else. For example, a causual browser of your Salesforce being one of your admin staff shouldn't get the same SLA as someone in the operations department trying to update an order-inventory system. Ofcourse you wouldn't do app QoS for a single user but do it more based on their roles in enterprise.  BTW, there are startups working on providing Cloud services for Video and Voice.
 
There are two parts to the cloud that are interesting - 
(1) front end - the box where users connect to the cloud.  For front end content aware switches would play a key role
(2) back end - Storage side where FCOE, FC and Infiniband switches get deployed.
 
Everything in the middle should just be connectivity points. Fat or no so fat connectivity points would depend on your pockets being fat or no so fat.

 

> Subject: RE: "Cloud Networking" defined in Arista Networks new CEO Jayshree Ullal's blog

> Date: Fri, 24 Oct 2008 16:18:06 -0700
> From: ksa...@cisco.com
> To: cloud-c...@googlegroups.com

[Sundeep Singatwaria]

 
Shreyas - Distribute software and configuration is more like using one box to download images to all other boxes in the cloud network so that you don't have to login to each box and download image on it. Same applies to the configuration. Almost all networking companies have it or are working on it.
 
What is stateless networking - can some one explain.

---

Date: Mon, 3 Nov 2008 12:19:38 -0800
From: shreyas...@yahoo.com
Subject: [ Cloud Computing ] Re: "Cloud Networking" defined in Arista Networks new CEO Jayshree Ullal's blog
To: cloud-c...@googlegroups.com

</HTML

</BLOCKQUOTE

[Ray Nugent]

better still, can someone explain statefull networking?

Ray

---

From: Sundeep Singatwaria <sundeep_s...@hotmail.com>
To: cloud-c...@googlegroups.com
Sent: Monday, November 3, 2008 1:53:52 PM

[mike.dvorkin]

This can all be solved a lot simpler if the problem is reduced to a traditional policy management problem. A configuration, an image, an application is merely a rendering of policy effects. Such policy effects need to be fully resolved at each point in the fabric.

A notion of centralized explicit management in the cloud is not achievable at cloud's scale.

Mike

---

From: Sundeep Singatwaria <sundeep_s...@hotmail.com>
To: cloud-c...@googlegroups.com
Sent: Monday, November 3, 2008 1:53:52 PM

[Sundeep Singatwaria]

 
Ray - Stateful networking imply different things to different people. If you look at content aware switches - stateful networking would mean
establishing a flow record for a user and then forwarding all the traffic based on flow record. The flow record could be as simple as a five tuple flow record used in firewalls today. This traffic would be from the inbound connectivity point of the cloud to the end point switch that connects to my storage.
 
Stateful networking from a pure switching routing world would mean establishing a flow state from start to the end-point routers so that a user traffic get deterministic per hop behaviour at each intermediate point in the network. That would be my outbound traffic from my enterprise to the inbound connectivity point in the cloud.
 
So my flow would have two stateful states
(1) A switching / routing state forwarding state from my enterprise to the inbound connectivity point of the cloud.
(2) A flow record state from inbound connectivity point in the cloud to the last hop switch that connects to my storage.
 
Stateful networking would thus provide determistic per hop behaviour for the flow end-to-end. It clouds were to be useful for my business, I would expect it to provide determistic behaviour much like the enterprise applications hosted in a Datacenter within the enterprise.
 
If I were to do stateless networking, how do I get determistic behaviour for my apps. Well, that brings me back to my original question - What is stateless networking ?
 
Sundeep

---

Date: Mon, 3 Nov 2008 14:14:58 -0800
From: rnu...@yahoo.com

[shreyas shah]

Switches typically remembers the flows that are set (statically) and hence provide the QOS required per flow/application.

 

In the Stateless networking, it is changed dynamically (Quasi-static) based on the customer traffic needs and these changes have to be propagated to all the networking devices in the infrastructure in minimal time to be effective.

 

EOS will allow reconfiguration of the switches and hence QOS will be enforced on these flows.

[Sundeep Singatwaria]

 
Shreyas - Aren't they more stateful than the switches we know today. reconfiguration would require one to retune the QOS params to enforce QOS.

 

---

Date: Tue, 11 Nov 2008 11:40:20 -0800

</HTML

<BR

[shreyas shah]

Per flow basis, switches are more stateful but they lack statelessness to be reconfigured on the fly and enforce QOS/SLA per customer flow.

 

Arista switches will provide following so it can be used as part of Cloud Network infrastructure,

1. Customization of the switches

2. Reconfig on the fly

 

Arista will provide the interfaces to DDC (Dynamic data center) management tools.