NVM Express over Fabrics


Any technologist who’s read, let alone used NVM Express (NVMe), is pretty enthusiastic about it’s capabilities and if it was not for availability and financial restrictions we’d all have at least a couple in our home systems and labs. It seems to succeed very well in making sure the host can keep up with the performance (low latencies, high throughput)) delivered by SSD drives than our current interfaces.

This means that many are very happy with future visions on how PCIe will dislodge SAS/SATA as the preferred SSD interface. This might seem feasible for local storage right now but how to deal with this in an actual storage array, what if we want to size this to a larger scale? There are no “PCIe JBODS”. So what does one do? Well, how did we do it in the past with FC? We created a fabric. Below we see several local & remote NVMe architectures even hybrid ones with traditional SAS.

image

That’s exactly what NVM Express Inc. is doing, creating the specs for a fabric. This holds the promise to achieve superior results due to the elimination of SCSI translation which reduces latency significantly by delivering NVMe end to end. Not only that but we also see the following efforts in the NVM Express Specification 1.2 to give it enterprise grade capabilities beyond pure performance.

  • Enhanced status reporting
  • Expanded capabilities including live firmware updates

There have been some early demos of NVMe over Fabrics mainly focusing on the “remote” performance. While local NVMe SSDs have the edge on absolute IOPS the difference with NVMe over a fabric is not significant. The reduction in latency is measured in < 10 µs,so that’s good news. The fabric leverages RDMA (yes, yet another reason that my time spending with this technology has been a useful investment). This can be Infiniband, RoCE or iWarp. There’s also the new kid on the block “Intel Omni Scale”  (even if their early demo used iWARP). There’s also a Mellanox RoCE demo.

image

Now with NVMe SSD disk speeds it seems that the writing is on the wall that ever better fabric performance will be needed to support the tremendous throughput this evolution of storage can deliver. RDMA seems poised for success in regards to this. Now, yes, strictly speaking the NVMe traffic does not require RDMA but let’s just say I don’t see anyone building it without. I also think this means even iWarp fabrics will use DCB (PFC) to make sure we have a lossless network. The amount of traffics will be immense and why not optimize for the best possible performance? I hold the opinion this is beneficial for east-west traffic today in larger environments, especially when in converged networks. Unless the Intel® Omni-Path Architecture blows everyone else away that is Smile. Too early to tell.

Now does this dictate the total and absolute obsolescence of iSCSI and FC? No. There is no reason why a NVMe Fabrics storage solution cannot offer storage to hosts via FC, iSCSI, SMB 3, NFS, FCoE, … They, potentially could even offer iWarp, RoCE or Infiniband to the hosts so you won’t lose your prior investments or get locked into one. I have no magic ball so I cannot tell you if this will happen. What I do now that when it comes to MPIO versus multichannel for load balancing and even failover and recovery, multichannel sometimes does a (far?) superior job in my honest opinion especially with older hypervisors even when the hypervisor uses separate sessions per virtual machine to achieve better load balancing over iSCSI or the like. Anyway, I digress. One thing I do know is that I’ll keep a keen eye on what Microsoft is doing in this space, especially in regards to Windows Server 2016. It’s time to up the level on scalability & support for newer technologies once again.

Advertisements

DCB ETS Demo with SMB Direct over RoCE (RDMA)


It’s time to demonstrate ETS in action! There is a quick video on ETS on Vimeo to show what it look like.

I’m using Mellanox ConnectX-3 ethernet cards, in 2 node DELL PowerEdge R720 Hyper- cluster lab. We’ve configured the two ports for SMB Direct & set live migration to leverage them both over SMB Direct. For the purpose of this demo we’ll generate non RDMA over RoCE (TCP/IP) traffic over these two 10Gbps ports to simulate a problematic scenario where all bandwidth is already being used and to see how Enhanced Transmission Selection (ETS) will help in this scenario.  I have done this with DELL Force 10, PowerConnect 8100, N4000 series or a mix of both. This particular demo was leveraging PC8132Fs. I use what’s available to me in a lab at the time of writing.

To achieve the network load this we leverage ntttcp.exe to generate the non RDMA TCP/IP traffic. Using the Mellanox QoS counters we visualize this. In blue you see the sending traffic from node A, in red the receiving traffic on Node B. Note that this traffic is tagged with priority 1. We tag SMB Direct traffic with priority 4.

image

You can see that both Mellanox cards are running at full bandwidth, 2* 10Gbps from node A to node B and it’s all none RDMA traffic. Also note that I’m hitting all 16 physical cores (hyper threading is enabled). By doing so I avoid being bottlenecked by a singe core as in contrast to RDMA traffic there’s no huge CPU offload going on here.image

As these are the cards I have assigned to use for live migration (depending on the setup also  CSV or SOFS traffic) over SMB Direct you’ll see that the competition for bandwidth will be fierce if we don’t have a mechanism to guide this to a desired outcome. That’s exactly what we leverage DCB with PFC and ETS for.

So let’s kick off live migration of 4 virtual machines with 10GB of memory each. That should take about 20 seconds on 2 * 10Gbps cards. We first live migrate them form node B to Node A. That’s in the reverse direction of where we are sending TCP/IP traffic. You see 10Gbps being used all over and this is expected.

image

Remember that the network is full duplex. That means that you can send at 10Gbps (TCP/IP from node A to node B, RDMA from node B to A and vice versa) and receive at 10Gbps on a port. Actually if the backplane of the switch is powerful enough you can do so on all ports. So this is normal. Node A is sending TCP/IP traffic to node B at line speed and Node B is sending SMB Direct traffic to node A (the live migration) at line speed.

But what if we live migrate over SMB Direct in the same direction as the TCP/IP traffic is going, from node A to node B? Well have a look. To me this looks awesome.

image

ETS kicks in immediately. We configure the minimum bandwidth for SMB Direct Traffic to be 90%. Anything left after that (10%) is given to other traffic, in this demo the TCP/IP traffic we generated. As priority 4 tagged RoCE traffic is also configured to be lossless with PFC you don’t have to worry about dropping packets under contention. Now think about this and how you can steer your traffic behavior at times when the resources need to be divided amongst competing workloads.

I hope you now have a better idea on why QoS is useful, how it works and that it indeed does work. While I have taken the opportunity to demonstrate this with SMB Direct over RoCE I’d like to stress that QoS is not just about RoCE where it’s  “mandatory” due to the fact it requires at least PFC. It’s a very much a needed tool that’s very beneficial in any converged scenario and that the optional ETS might be a very good idea, depending on your environment.

Again, to get you a better idea, here’s a short, quick video on ETS on Vimeo.

DCB PFC Demo with SMB Direct over RoCE (RDMA)


In this blog post we’ll demo Priority Flow control. We’re using the demo comfit as described in SMB Direct over RoCE Demo – Hosts & Switches Configuration Example

There is also a quick video to illustrate all this on Vimeo. It’s not training course grade I know, but my time to put into these is limited.

I’m using Mellanox ConnectX-3 ethernet cards, in 2 node DELL PowerEdge R720 Hyper- cluster lab. We’ve configured the two ports for SMB Direct & set live migration to leverage them both over SMB Direct. For that purpose we tagged SMB Direct traffic with priority 4 and all other traffic with priority 1. We only made priority lossless as that’s required for RoCE and the other traffic will deal with not being lossless by virtue of being TCP/IP.

Priority Flow Control is about making traffic lossless. Well some traffic. While we’d love to live by Queens lyrics “I want it all, I want it all and I want it now” we are limited. If not so by our budgets, than most certainly by the laws of physics. To make sure we all understand what PFC does here’s a quick reminder: It tells the sending party to stop sending packets, i.e. pause a moment (in our case SMB Direct traffic) to make sure we can handle the traffic without dropping packets. As RoCE is for all practical purposes Infiniband over Ethernet and is not TCP/IP, so you don’t have the benefits of your protocol dealing with dropped packets, retransmission … meaning the fabric has to be lossless*. So no it DOES NOT tell non priority traffic to slow down or stop. If you need to tell other traffic to take a hike, you’re in ETS country 🙂

* If any switch vendor tells you to not bother with DCB and just build (read buy their switches = $$$$$) a lossless fabric (does that exist?) and rely on the brute force quality of their products to have a lossless experience … could be an interesting experiment Smile.

Note: To even be able to start SMB Direct SMB Multichannel must be enabled as this is the mechanism used to identify RDMA capabilities after which a RDMA connection is attempted. If this fails you’ll fall back to SMB Multichannel. So you will have ,network connectivity.

You want RDMA to work and be lossless. To visualize this we can turn to the switch where we leverage the counter statistics to see PFC frames being send or transmitted. A lab example from a DELL PowerConnect 8100/N4000 series below.

image

To verify that RDMA is working as it should we should also leverage the Mellanox Adapter Diagnostic and native Windows RDMA Activity counters. First of all make sure RDMA is working properly. Basically you want the error counters to be zero and stay that way.

Mellanox wise these must remain at zero (or not climb after you got it right):

  • Responder CQE Errors
  • Responder Duplicate Request Received
  • Responder Out-Of-Order Sequence Received
  • … there’s lots of them …

image

Windows RDMA Activity wise these should be zero (or not climb after you got it right):

  • RDMA completion Queue Errors
  • RDMA connection Errors
  • RDMA Failed connection attempts

image

The event logs are also your friend as issues will log entries to look out for like

PowerShell is your friend (adapt severity levels according to your need!)

Get-WinEvent -ListLog “*SMB*” | Get-WinEvent | ? { $_.Level -lt 4 -and $_. Message -like “*RDMA*” } | FL LogName, Id, TimeCreated, Level, Message

Entries like this are clear enough, it ain’t working!

The network connection failed.
Error: The I/O request was canceled.
Connection type: Rdma
Guidance:
This indicates a problem with the underlying network or transport, such as with TCP/IP, and not with SMB. A firewall that blocks port 445 or 5445 can also cause this issue.
 
RDMA interfaces are available but the client failed to connect to the server over RDMA transport.
Guidance:
Both client and server have RDMA (SMB Direct) adaptors but there was a problem with the connection and the client had to fall back to using TCP/IP SMB (non-RDMA).

 

To view PFC action in Windows we rely on the Mellanox Adapter QoS Counters

image

Below you’ll see the number of  pause frames being sent & received on each port. Click on the image to enlarge.

image

An important note trying to make sense of it all: … pauze and receive frames are sent and received hop to hop. So if you see a pause frame being sent on a server NIC port you should see them being received on the switch port and not on it’s windows target you are live migrating from. The 4 pause frames sent in the screenshot above are received by the switchport as you can see from the PFC Stats for that port.

image

People, if you don’t see errors in the error counters and event viewer that’s good. If you see the PFC Pause frame counters move up a bit that’s (unless excessive) also good and normal, that PFC doing it’s job making sure the traffic is lossless. If they are zero and stay zero for ever you did not buy a lossless fabric that doesn’t need DCB, it’s more likely you DCB/PFC is not working Winking smile and you do not have a lossless fabric at all. The counters are cumulative over time so they don’t reset to zero bar resetting the NIC or a reboot.

image

When testing feel free to generate lots of traffic all over the place on the involved ports & switches this helps with seeing all this in action and verifying RDMA/PFC works as it should. I like to use ntttcp.exe to generate traffic, the most recent version will let you really put a load on 10GBps and higher NICs. Hammer that network as hard as you can Winking smile.

Again a simple video to illustrate this on Vimeo.