Network Connectivity for Clones

This document presents the strategy to ensure continued network connectivity for multiple clones created from a single Firecracker microVM snapshot. This document also provides an overview of the scalability benchmarks we performed.

Setup

There are two things which prevent network connectivity from resuming out-of-the-box for clones created from the same snapshot: Firecracker currently saves and attempts to restore network devices using the initially configured TAP names, and each guest will be resumed with the same network configuration, most importantly with the same IP address(es). To work around the former, each clone should be started within a separate network namespace (we can have multiple TAP interfaces with the same name, as long as they reside in distinct network namespaces). The latter can be mitigated by leveraging iptables SNAT and DNAT support. We choose a clone address (CA) for each clone, which is the new address that’s going to represent the guest, and make it so all packets leaving the VM have their source address rewritten to CA, and all incoming packets with the destination address equal to CA have it rewritten to the IP address configured inside the guest (which remains the same for all clones). Each individual clone continues to believe it’s using the original address, but outside the VM packets are assigned a different one for every clone.

Let’s have a more detailed look at this approach. We assume each VM has a single network interface attached. If multiple interfaces with full connectivity are required, we simply repeat the relevant parts of this process for each additional interface. A typical setup right before taking a snapshot involves having a VM with a network interface backed by a TAP device (named vmtap0, for example) with an IP address (referred to as the TAP IP address, for example 192.168.241.1/29), and an IP address configured inside the guest for the corresponding virtio device (referred to as the guest IP address, for example 192.168.241.2/29).

Network namespaces

Attempting to restore multiple clones from the same snapshot faces the problem of every single one of them attempting to use a TAP device with the original name, which is not possible by default. Therefore, we need to start each clone in a separate network namespace. This is already possible using the netns jailer parameter, described in the documentation. The specified namespace must already exist, so we have to create it first using

sudo ip netns add fc0

(where fc0 is the name of the network namespace we plan to use for this specific clone - clone0). A new network namespace is initially empty, so we also have to create a new tap interface within using

sudo ip netns exec fc0 ip tuntap add name vmtap0 mode tap

The ip netns exec <ns_name> <command> allows us to execute command in the context of the specified network namespace (in the previous case, the secondary command creates a new tap interface). Next we configure the new TAP interface to match the expectations of the snapshotted guest by running

sudo ip netns exec fc0 ip addr add 192.168.241.1/29 dev vmtap0
sudo ip netns exec fc0 ip link set vmtap0 up

At this point we can start multiple clones, each in its separate namespace, but they won’t have connectivity to the rest of the host, only the respective TAP interfaces. However, interaction over the network is still possible; for example we can connect over ssh to clone0 using

sudo ip netns exec fc0 ssh root@192.168.241.2

veth interfaces to connect the network namespaces

In order to obtain full connectivity we have to begin by connecting the network namespace to the rest of the host, and then solving the “same guest IP” problem. The former requires the use of veth pairs - virtual interfaces that are link-local to each other (any packet sent through one end of the pair is immediately received on the other, and the other way around). One end resides inside the network namespace, while the other is moved into the parent namespace (the host global namespace in this case), and packets flow in or out according to the network configuration. We have to pick IP addresses for both ends of the veth pair. For clone index idx, let’s use 10.<idx / 30>.<(idx % 30) * 8>.1/24 for the endpoint residing in the host namespace, and the same address ending with 2 for the other end which remains inside the clone's namespace. Thus, for clone 0 the former is 10.0.0.1 and the latter 10.0.0.2.

The first endpoint must have an unique name on the host, for example chosen as veth(idx + 1) (so veth1 for clone 0). To create and setup the veth pair, we use the following commands (for namespace fc0):

# create the veth pair inside the namespace
sudo ip netns exec fc0 ip link add veth1 type veth peer name veth0
# move veth1 to the global host namespace
sudo ip netns exec fc0 ip link set veth1 netns 1

sudo ip netns exec fc0 ip addr add 10.0.0.2/24 dev veth0
sudo ip netns exec fc0 ip link set dev veth0 up

sudo ip addr add 10.0.0.1/24 dev veth1
sudo ip link set dev veth1 up

# designate the outer end as default gateway for packets leaving the namespace
sudo ip netns exec fc0 ip route add default via 10.0.0.1

iptables rules for end-to-end connectivity

The last step involves adding the iptables rules which change the source/destination IP address of packets on the fly (thus allowing all clones to have the same internal IP). We need to choose a clone address, which is unique on the host for each VM. In the demo, we use 192.168.<idx / 30>.<(idx % 30) * 8 + 3> (which is 192.168.0.3 for clone 0):

# for packets that leave the namespace and have the source IP address of the
# original guest, rewrite the source address to clone address 192.168.0.3
sudo ip netns exec fc0 iptables -t nat -A POSTROUTING -o veth0 \
-s 192.168.241.2 -j SNAT --to 192.168.0.3

# do the reverse operation; rewrites the destination address of packets
# heading towards the clone address to 192.168.241.2
sudo ip netns exec fc0 iptables -t nat -A PREROUTING -i veth0 \
-d 192.168.0.3 -j DNAT —to 192.168.241.2

# (adds a route on the host for the clone address)
sudo ip route add 192.168.0.3 via 10.0.0.2

Full connectivity to/from the clone should be present at this point.

To make sure the guest also adjusts to the new environment, you can explicitly clear the ARP/neighbour table in the guest:

ip -family inet neigh flush any
ip -family inet6 neigh flush any

Otherwise, packets originating from the guest might be using old Link Layer Address for up to arp cache timeout seconds. After said timeout period, connectivity will work both ways even without an explicit flush.

Scalability evaluation

We ran synthetic tests to determine the impact of the addtional iptables rules and namespaces on network performance. We compare the case where each VM runs as regular Firecracker (gets assigned a TAP interface and a unique IP address in the global namespace) versus the setup with a separate network namespace for each VM (together with the veth pair and additional rules). We refer to the former as the basic case, while the latter is the ns case. We measure latency with the ping command and throughput with iperf.

The experiments, ran on an Amazon AWS m5d.metal EC2 instace, go as follows:

• Set up 3000 network resource slots (different TAP interfaces for the basic case, and namespaces + everything else for ns). This is mainly to account for any difference the setup itself might make, even if there are not as many active endpoints at any given time.
• Start 1000 Firecracker VMs, and pick N < 1000 as the number of active VMs that are going to generate network traffic. For ping experiments, we ping each active VM from the host every 500ms for 30 seconds. For iperf experiments, we measure the average bandwidth of connections from the host to every active VM lasting 40 seconds. There is one separate client process per VM.
• When N = 100, in the basic case we get average latencies of 0.315 ms (0.160/0.430 min/max) for ping, and an average throughput of 2.25 Gbps (1.62/3.21 min/max) per VM for iperf. In the ns case, the ping results are bumped higher by around 10-20 us, while the iperf results are virtually the same on average, with a higher minimum (1.73 Gbps) and a lower maximum (3 Gbps).
• When N = 1000, we start facing desynchronizations caused by difficulties in starting (and thus finishing) the client processes all at the same time, which creates a wider distribution of results. In the basic case, the average latency for ping experiments has gone down to 0.305 ms, the minimum decreased to 0.155 ms, but the maximum increased to 0.640 ms. The average iperf per VM throughput is around 440 Mbps (110/3936 min/max). In the ns case, average ping latency is now 0.318 ms (0.170/0.650 min/max). For iperf, the average throughput is very close to basic at ~430 Mbps, while the minimum and maximum values are lower at 85/3803 Mbps.

The above measurements give a significant degree of confidence in the scalability of the solution (subject to repeating for different values of the experimental parameters, if necessary). The increase in latency is almost negligible considering usual end-to-end delays. The lower minimum throughput from the iperf measurements might be significant, but only if that magnitude of concurrent, data-intensive transfers is likely. Moreover, the basic measurements are close to an absolute upper bound.