Swapping over NBD is something that is technically possible but not
often advised. While there are number of guides on the internet
on how to configure it and nbd-client supports a -swap switch to
"prevent deadlocks", the fact of the matter is a machine using NBD
for swap can be locked up within minutes if swap is used intensively.
The problem is that network block devices do not use mempools like
normal block devices do. As the host cannot control where they receive
packets from, they cannot reliably work out in advance how much memory
they might need.
Some years ago, Peter Ziljstra developed a series of patches that
supported swap over an NFS that some distributions are carrying in
their kernels. This patch series borrows very heavily from Peter's work
to support swapping over NBD (the relatively straight-forward case)
and uses throttling instead of dynamically resized memory reserves
so the series is not too unwieldy for review.
Patch 1 serialises access to min_free_kbytes. It's not strictly needed
by this series but as the series cares about watermarks in
general, it's a harmless fix. It could be merged independently.
Patch 2 adds knowledge of the PFMEMALLOC reserves to SLAB and SLUB to
preserve access to pages allocated under low memory situations
to callers that are freeying memory.
Patch 3 introduces __GFP_MEMALLOC to allow access to the PFMEMALLOC
reserves without setting PFMEMALLOC.
Patch 4 opens the possibility for softirqs to use PFMEMALLOC reserves
for later use by network packet processing.
Patch 5 ignores memory policies when ALLOC_NO_WATERMARKS is set.
Patches 6-9 allows network processing to use PFMEMALLOC reserves when
the socket has been marked as being used by the VM to clean
pages. If packets are received and stored in pages that were
allocated under low-memory situations and are unrelated to
the VM, the packets are dropped.
Patch 10 is a micro-optimisation to avoid a function call in the
common case.
Patch 11 tags NBD sockets as being SOCK_MEMALLOC so they can use
PFMEMALLOC if necessary.
Patch 12 notes that it is still possible for the PFMEMALLOC reserve
to be depleted. To prevent this, direct reclaimers get
throttled on a waitqueue if 50% of the PFMEMALLOC reserves are
depleted. It is expected that kswapd and the direct reclaimers
already running will clean enough pages for the low watermark
to be reached and the throttled processes are woken up.
Some basic performance testing was run using kernel builds, netperf
on loopback for UDP and TCP, hackbench (pipes and sockets), iozone
and sysbench. Each of them were expected to use the sl*b allocators
reasonably heavily but there did not appear to be significant
performance variances. Here is the results from netperf using
slab as an example
NETPERF UDP
netperf-udp udp-swapnbd
vanilla-slab v1r17-slab
64 178.06 ( 0.00%)* 189.46 ( 6.02%)
1.02% 1.00%
128 355.06 ( 0.00%) 370.75 ( 4.23%)
256 662.47 ( 0.00%) 721.62 ( 8.20%)
1024 2229.39 ( 0.00%) 2567.04 (13.15%)
2048 3974.20 ( 0.00%) 4114.70 ( 3.41%)
3312 5619.89 ( 0.00%) 5800.09 ( 3.11%)
4096 6460.45 ( 0.00%) 6702.45 ( 3.61%)
8192 9580.24 ( 0.00%) 9927.97 ( 3.50%)
16384 13259.14 ( 0.00%) 13493.88 ( 1.74%)
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 2960.17 2540.14
Total Elapsed Time (seconds) 3554.10 3050.10
NETPERF TCP
netperf-tcp tcp-swapnbd
vanilla-slab v1r17-slab
64 1230.29 ( 0.00%) 1273.17 ( 3.37%)
128 2309.97 ( 0.00%) 2375.22 ( 2.75%)
256 3659.32 ( 0.00%) 3704.87 ( 1.23%)
1024 7267.80 ( 0.00%) 7251.02 (-0.23%)
2048 8358.26 ( 0.00%) 8204.74 (-1.87%)
3312 8631.07 ( 0.00%) 8637.62 ( 0.08%)
4096 8770.95 ( 0.00%) 8704.08 (-0.77%)
8192 9749.33 ( 0.00%) 9769.06 ( 0.20%)
16384 11151.71 ( 0.00%) 11135.32 (-0.15%)
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 1245.04 1619.89
Total Elapsed Time (seconds) 1250.66 1622.18
Here is the equivalent test for SLUB
NETPERF UDP
netperf-udp udp-swapnbd
vanilla-slub v1r17-slub
64 180.83 ( 0.00%) 183.68 ( 1.55%)
128 357.29 ( 0.00%) 367.11 ( 2.67%)
256 679.64 ( 0.00%)* 724.03 ( 6.13%)
1.15% 1.00%
1024 2343.40 ( 0.00%)* 2610.63 (10.24%)
1.68% 1.00%
2048 3971.53 ( 0.00%) 4102.21 ( 3.19%)*
1.00% 1.40%
3312 5677.04 ( 0.00%) 5748.69 ( 1.25%)
4096 6436.75 ( 0.00%) 6549.41 ( 1.72%)
8192 9698.56 ( 0.00%) 9808.84 ( 1.12%)
16384 13337.06 ( 0.00%) 13404.38 ( 0.50%)
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 2880.15 2180.13
Total Elapsed Time (seconds) 3458.10 2618.09
NETPERF TCP
netperf-tcp tcp-swapnbd
vanilla-slub v1r17-slub
64 1256.79 ( 0.00%) 1287.32 ( 2.37%)
128 2308.71 ( 0.00%) 2371.09 ( 2.63%)
256 3672.03 ( 0.00%) 3771.05 ( 2.63%)
1024 7245.08 ( 0.00%) 7261.60 ( 0.23%)
2048 8315.17 ( 0.00%) 8244.14 (-0.86%)
3312 8611.43 ( 0.00%) 8616.90 ( 0.06%)
4096 8711.64 ( 0.00%) 8695.97 (-0.18%)
8192 9795.71 ( 0.00%) 9774.11 (-0.22%)
16384 11145.48 ( 0.00%) 11225.70 ( 0.71%)
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 1345.05 1425.06
Total Elapsed Time (seconds) 1350.61 1430.66
Time to completion varied a lot but this can happen with netperf as
it tries to find results within a sufficiently high confidence. I
wouldn't read too much into the performance gains of netperf-udp
as it can sometimes be affected by code just shuffling around for
whatever reason.
For testing swap-over-NBD, a machine was booted with 2G of RAM with a
swapfile backed by NBD. 16*NUM_CPU processes were started that create
anonymous memory mappings and read them linearly in a loop. The total
size of the mappings were 4*PHYSICAL_MEMORY to use swap heavily under
memory pressure. Without the patches, the machine locks up within
minutes and runs to completion with them applied.
Comments?
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>