Skip to content

xzpeter/mig_mon

Repository files navigation

About

mig_mon is the short form of Migration Monitor. It's a set of tools for VM migration testing and debugging.

Features

mig_mon provides a few sub-commands to use.

VM Live Migration Network Emulator

This sub-tool can be used to emulate live migration TCP streams.

There're two types of live migration: (1) precopy (2) postcopy. This tool can emulate (1) or (2) or (1+2) case by specifying different '-t' parameters:

  • Enable precopy only: it emulates a TCP_STREAM workload from src->dst
  • Enable postcopy only: it emulates a TCP_RR workload from dst->src
  • Enable both: it emulates the above TCP_STREAM+TCP_RR on the same socket

For precopy stream, it's the bandwidth that matters. The bandwidth information will be dumped per-second on src VM.

For postcopy stream, it's the latency that matters. The average/maximum latency value of page requests will be dumped per-second on dst VM.

This sub-command has below parameters:

./mig_mon vm [options...]
  -d:    Emulate a dst VM
  -h:    Dump help message
  -H:    Specify dst VM IP (required for -s)
  -s:    Emulate a src VM
  -S:    Specify size of the VM (GB)
  -t:    Specify tests (precopy, postcopy)

Example usage:

To start the (emulated) destination VM, one can run this on dest host:

./mig_mon vm -d

Then, to start a src VM emulation and start both live migration streams, one can run this command on src host:

./mig_mon vm -s -H $DEST_IP -t precopy -t postcopy

Specifying both '-t' will just enable both migration streams.

Memory Dirty

Sub-command "mm_dirty" can generate a constant dirty workload in the system.

./mig_mon mm_dirty [options...]
  -h:    Dump help message for mm_dirty sub-cmd
  -m:    Memory size in MB (default: 512)
  -r:    Dirty rate in MB/s (default: unlimited)
  -p:    Work pattern: "sequential", "random", or "once"
         (default: "sequential")
  -L:    Record and report memory access latencies
  -P:    Page size: "2m" or "1g" for huge pages

To generate a random dirty workload of 500MB/s upon 2GB memory range, we can use:

./mig_mon mm_dirty -m 2000 -r 500 -p random

The dirty workload will always dirty pages in 4K page size (even if huge pages are used) because normally hypervisor will trap dirty in small page size always.

Pre-heat will be done before starting the real workload.

Memory Access Latency Measurement

mm_dirty can also support measuring memory access latencies during writing to memory. It's mostly useful when e.g. there's a potential reason for high memory access latency (e.g. the VM is during a postcopy live migration), then we can get a distribution of memory access latencies for the whole process.

To record and report memory access latencies, simply attach parameter -L to the mm_dirty command. Below is an example to start sequential writes upon 16GB memory, measure / report memory access latencies:

./mig_mon mm_dirty -m 16G -L

The result (on a bare metal host) can look like this:

        1 (us): 23372101
        2 (us): 2399961
        4 (us): 2168
        8 (us): 1454
       16 (us): 76
       32 (us): 5
       64 (us): 0
      128 (us): 0
      256 (us): 0
      512 (us): 0
     1024 (us): 0
     2048 (us): 0
     4096 (us): 0
     8192 (us): 0
    16384 (us): 0
    32768 (us): 0
    65536 (us): 0
   131072 (us): 0
   262144 (us): 0
   524288 (us): 0
  1048576 (us): 0

Note that there're 21 buckets, each of the bucket is a power-of-2. For example, the number showed in bucket 8 (us) means there are 1454 memory accesses that took no more than 8 microseconds to finish (but larger than 4us or it'll have fallen into the previous bucket). Same applies to the rest buckets.

Here only the last bucket is special: anything bigger than 1sec will be put there.

Network Downtime Measurement

Sub-command "server_rr/client" can be used to measure guest OS network downtime during migration. To use it, we can first start the UDP echo server in the guest using:

./mig_mon server_rr

Then from outside the guest, we can start the client trying to send a packet for constant interval (e.g. 50ms) and waiting for a response:

./mig_mon client $GUEST_IP 50 $LOG

The client side will record the latency of each packet received, recording spikes into $LOG and also show the maximum latency detected.

About

VM migration related toolset

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages