Bi-modal streaming replication throughput

[Alexis Lê-Quôc <alq@xxxxxxxxxxxxx>]

Hi,

I have been puzzled by very different replication performance (meaning 50-100x slower) between identical replicas (both in “hardware” and configuration) once the amount of data to replicate increases. I’ve gone down a number of dead ends and am missing something

(

likely obvious

)

that I hope folks with a deeper knowledge can point out. I’ve tried to boil down the data need to describe the issue to a minimum.

 Thanks for taking the time to read and for any ideas you can share.

# The setup

We run

a cluster of

large, SSD-backed, i3.16xl (64 cores visible to Linux, ~500GB of RAM, with 8GB of shared_buffers, fast NVMe drives) nodes

, each

running PG 9.3

on linux

in a vanilla streaming asynchronous replication setup: 1 primary node, 1 replica designated for failover (left alone) and 6 read replicas, taking queries.

Under normal circumstances this is working exactly as planned but when I dial up the number of INSERTs on the primary to ~10k rows per second, or roughly 50MB of data per second (not enough to saturate the network between nodes)

, read replicas falls hopelessly and consistently behind until read traffic is diverted away

. The INSERTs themselves are fairly straightforward: a 20-bytea checksum is computed off-node

and used as a unicity constraint at insert time. Each record is 4,500 bytes wide on average.

H

ere’s the table where inserts happen.

                                                  Table “T”
     Column     |            Type             |               Modifiers                | Storage  |
----------------+-----------------------------+----------------------------------------+----------+
 key            | bigint                      | not null default T.next_key()      

   

| plain    |
 a              | integer                     | not null                               | plain    |
 b              | integer                     |                                        | plain    |
 c              | text                        |                                        | extended |
 d              | text                        |                                        | extended |
 e              | text[]                      |                                        | extended |
 f              | integer                     | not null                               | plain    |
 created        | timestamp without time zone | not null default now()                 | plain    |
 cksum          | bytea                       | not null                               | extended |
Indexes:
    “T_pkey" PRIMARY KEY, btree (key)
    “T_cksum” UNIQUE, btree (cksum)
    “T_created_idx" btree (created)
    “T_full_idx" btree (a, b, c, d, e)
    “T_a_idx" btree (a)


# The symptoms

Once the primary starts to process INSERTs to the tune of 10k/s (roughly

5

0MB/s or 150GB/h), replication throughput becomes bi-modal

 within minutes.

1. We see read replicas fall behind and we can measure their replication throughput to be

consistently

1-2% of what the primary is sustaining, by measuring the replication delay (in second) every second. We quickly get

that metric

to 0.98-0.99 (1 means that replication is completely stuck

as it falls behind by one second every second

). CPU, memory

, I/O

(per core iowait)

or network

(throughput)

as a whole resource are not

visibly

maxed out

.

2. If we stop incoming queries from one of the replicas, we see it catch up at 2x insert throughput (roughly 80MB/s or 300GB/h) as it is cutting through the backlog. A perf sample shows a good chunk of time spent in `mdnblocks`. I/O wait remains

at

a few %

(2-10) of cpu cycles. If you can open the attached screenshot you can see the lag going down on each replica as soon as we stop sending reads at it.

In both cases the recovery process maxes out 1 core

as expected

.

# The question

What surprised me is the bi-modal nature of throughput without gradual degradation

or a very clear indication of the contentious resource (I/O? Buffer access?)

.

The bi-modal throughput

 would be consistent with replication being

effectively

scheduled to run

at full speed

1% or 2% of the time (the rest being allocated to queries) but I have not found something in the documentation or in the code that 

supports that view.

Is this the right way to think about what’s observed?

If not, what could be a good next hypothesis to test?

# References

Here are some settings that may help and a perf profile of a recovery process that runs without any competing read traffic processing the INSERT backlog (I don't unfortunately have the same profile on a lagging read replica).

             name             |  setting  
------------------------------+-----------
 max_wal_senders              | 299
 max_wal_size                 | 10240
 min_wal_size                 | 5
 wal_block_size               | 8192
 wal_buffers                  | 2048
 wal_compression              | off
 wal_keep_segments            | 0
 wal_level                    | replica
 wal_log_hints                | off
 wal_receiver_status_interval | 10
 wal_receiver_timeout         | 60000
 wal_retrieve_retry_interval  | 5000
 wal_segment_size             | 2048
 wal_sender_timeout           | 60000
 wal_sync_method              | fdatasync
 wal_writer_delay             | 200
 wal_writer_flush_after       | 128
shared_buffers           | 1048576
work_mem                        | 32768
maintenance_work_mem            | 2097152

recovery process sampled at 997Hz on a lagging replica without read traffic.

Samples: 9K of event 'cycles', Event count (approx.): 25040027878

  Children      Self  Command   Shared Object      Symbol

+   97.81%     0.44%  postgres  postgres           [.] StartupXLOG

+   82.41%     0.00%  postgres  postgres           [.] StartupProcessMain

+   82.41%     0.00%  postgres  postgres           [.] AuxiliaryProcessMain

+   82.41%     0.00%  postgres  postgres           [.] 0xffffaa514b8004dd

+   82.41%     0.00%  postgres  postgres           [.] PostmasterMain

+   82.41%     0.00%  postgres  postgres           [.] main

+   82.41%     0.00%  postgres  libc-2.23.so       [.] __libc_start_main

+   82.41%     0.00%  postgres  [unknown]          [k] 0x3bb6258d4c544155

+   50.41%     0.09%  postgres  postgres           [.] XLogReadBufferExtended

+   40.14%     0.70%  postgres  postgres           [.] XLogReadRecord

+   39.92%     0.00%  postgres  postgres           [.] 0xffffaa514b69524e

+   30.25%    26.78%  postgres  postgres           [.] mdnblocks

+   27.35%     0.00%  postgres  postgres           [.] heap_redo

+   26.23%     0.01%  postgres  postgres           [.] XLogReadBuffer

+   25.37%     0.05%  postgres  postgres           [.] btree_redo

+   22.49%     0.07%  postgres  postgres           [.] ReadBufferWithoutRelcache

+   18.72%     0.00%  postgres  postgres           [.] 0xffffaa514b6a2e6a

+   18.64%    18.64%  postgres  postgres           [.] 0x00000000000fde6a

+   18.10%     0.00%  postgres  postgres           [.] 0xffffaa514b65a867

+   15.80%     0.06%  postgres  [kernel.kallsyms]  [k] entry_SYSCALL_64_fastpath

+   13.16%     0.02%  postgres  postgres           [.] RestoreBackupBlock

+   12.90%     0.00%  postgres  postgres           [.] 0xffffaa514b675271

+   12.53%     0.00%  postgres  postgres           [.] 0xffffaa514b69270e

+   10.29%     0.00%  postgres  postgres           [.] 0xffffaa514b826672

+   10.00%     0.03%  postgres  libc-2.23.so       [.] write

+    9.91%     0.00%  postgres  postgres           [.] 0xffffaa514b823ffe

+    9.71%     0.00%  postgres  postgres           [.] mdwrite

+    9.45%     0.24%  postgres  libc-2.23.so       [.] read

+    9.25%     0.03%  postgres  [kernel.kallsyms]  [k] sys_write

+    9.15%     0.00%  postgres  [kernel.kallsyms]  [k] vfs_write

+    8.98%     0.01%  postgres  [kernel.kallsyms]  [k] new_sync_write

+    8.98%     0.00%  postgres  [kernel.kallsyms]  [k] __vfs_write

+    8.96%     0.03%  postgres  [xfs]              [k] xfs_file_write_iter

+    8.91%     0.08%  postgres  [xfs]              [k] xfs_file_buffered_aio_write

+    8.64%     0.00%  postgres  postgres           [.] 0xffffaa514b65ab10

+    7.87%     0.00%  postgres  postgres           [.] 0xffffaa514b6752d0

+    7.35%     0.04%  postgres  [kernel.kallsyms]  [k] generic_perform_write

+    5.77%     0.11%  postgres  libc-2.23.so       [.] lseek64

+    4.99%     0.00%  postgres  postgres           [.] 0xffffaa514b6a3347

+    4.80%     0.15%  postgres  [kernel.kallsyms]  [k] sys_read

+    4.74%     4.74%  postgres  [kernel.kallsyms]  [k] copy_user_enhanced_fast_string

Attachment: Screenshot 2018-08-14 13.16.52.png
Description: PNG image