On 31/10/2019 18:45, Paul Emmerich wrote:
On Fri, Oct 25, 2019 at 11:14 PM Maged Mokhtar <mmokhtar@xxxxxxxxxxx> wrote:
3. vmotion between Ceph datastore and an external datastore..this will be bad. This seems the case you are testing. It is bad because between 2 different storage systems (iqns are served on different targets), vaai xcopy cannot be used and vmware does its own stuff. It moves data using 64k block size, which gives low performance...to add some flavor, it does indeed use 32 threads, but unfortunately they use co-located addresses which does not work well in Ceph as they are hitting the same rbd object, which gets serialized due to pg locks, so you will not get any palatalization. Your speed will mostly be determined by a serial 64k, so with 1 ms write latency for ssd cluster, you will get around 64 MB/s..it will be slightly higher as the extra threads have some low effect.
Yes, vmotion is the worst IO pattern ever for a sequential copy.
However, the situation you are describing can be fixed with RBD
striping v2, just make Ceph switch to another object every 64kb, see
https://docs.ceph.com/docs/master/dev/file-striping/
I'm not sure about the state of striping v2 support in the kernel
module, last time I checked it wasn't supported. But
ceph-iscsi/tcmu-runner got quite good over the past year, I don't see
any point in still using the kernel data path for iscsi nowadays.
Paul
If case 3 (migration to/from Ceph to external stores) is important, then
maybe it is worth trying Veeam Quick Migration which supports live
migration. We have not tried it, but maybe this tool could be smarter
than vmware and use larger block sizes. They seem to have a free
edition, if you try it, i would recommend bumping the iSCSI MaxIoSizeKB
on the ESXi side as it helps with their backup tool.
I would not recommend using striping in this case, only very specific
cases benefit from it..but for general purpose vm workloads, striping
will probably be multiplying your queue depth and latencies with
negative effects, specially with such small sizes.
Ilya added striping to kernel rbd maybe in mainstream 4.17, we had sent
our own patches earlier but Ilya changes were more comprehensive. For
kernel rbd and librbd/tcmu, both have pros and cons, the kernel path is
used by SUSE, the main issue for us is the support for persistent
reservations which is required for setting up clustering on Windows.
/Maged
Note your esxtop does show 32 active ios under ACTV, the QUED of zero does is not the queue depth, but rather the "queued" io the ESX would suspend in case your active reaches the maximum by adapater ( 128 ).
This is just to clarify, if case 3 is not your primary concern than i would forget about it and benchmark 1 and 2 if they are relevant. Else, if 3 is important, i am not sure you can do much as it is happening within vmware..maybe there could be a way to map the external iqn to be served by the same target serving the Ceph iqn then there could be a chance the xcopy could be activated..Mike would probably know if this has any chance of working :)
/Maged
On 25/10/2019 22:01, Ryan wrote:
esxtop is showing a queue length of 0
Storage motion to ceph
DEVICE PATH/WORLD/PARTITION DQLEN WQLEN ACTV QUED %USD LOAD CMDS/s READS/s WRITES/s MBREAD/s MBWRTN/s DAVG/cmd KAVG/cmd GAVG/cmd QAVG/cmd
naa.6001405ec60d8b82342404d929fbbd03 - 128 - 32 0 25 0.25 1442.32 0.18 1440.50 0.00 89.78 21.32 0.01 21.34 0.01
Storage motion from ceph
DEVICE PATH/WORLD/PARTITION DQLEN WQLEN ACTV QUED %USD LOAD CMDS/s READS/s WRITES/s MBREAD/s MBWRTN/s DAVG/cmd KAVG/cmd GAVG/cmd QAVG/cmd
naa.6001405ec60d8b82342404d929fbbd03 - 128 - 32 0 25 0.25 4065.38 4064.83 0.36 253.52 0.00 7.57 0.01 7.58 0.00
I tried using fio like you mentioned but it was hanging with [r=0KiB/s,w=0KiB/s][r=0,w=0 IOPS] and the ETA kept climbing. I ended up using rbd bench on the ceph iscsi gateway. With a 64K write workload I'm seeing 400MB/s transfers.
rbd create test --size 100G --image-feature layering
rbd map test
mkfs.ext4 /dev/rbd/rbd/test
mount /dev/rbd/rbd/test test
rbd create testec --size 100G --image-feature layering --data-pool rbd_ec
rbd map testec
mkfs.ext4 /dev/rbd/rbd/testec
mount /dev/rbd/rbd/testec testec
[root@ceph-iscsi1 mnt]# rbd bench --image test --io-size 64K --io-type write --io-total 10G
bench type write io_size 65536 io_threads 16 bytes 10737418240 pattern sequential
SEC OPS OPS/SEC BYTES/SEC
1 6368 6377.59 417961796.64
2 12928 6462.27 423511630.71
3 19296 6420.18 420752986.78
4 26320 6585.61 431594792.67
5 33296 6662.37 436624891.04
6 40128 6754.67 442673957.25
7 46784 6765.75 443400452.26
8 53280 6809.02 446236110.93
9 60032 6739.67 441691068.73
10 66784 6698.91 439019550.77
11 73616 6690.88 438493253.66
12 80016 6654.35 436099640.00
13 85712 6485.07 425005611.11
14 91088 6202.49 406486113.46
15 96896 6021.17 394603137.62
16 102368 5741.19 376254347.24
17 107568 5501.57 360550910.38
18 113728 5603.17 367209502.58
19 120144 5820.48 381451245.32
20 126496 5917.60 387816078.53
21 132768 6089.71 399095466.00
22 139040 6306.98 413334431.09
23 145104 6276.42 411331743.63
24 151440 6256.67 410036891.68
25 157808 6261.12 410328554.98
26 163456 6140.03 402392725.36
elapsed: 26 ops: 163840 ops/sec: 6271.36 bytes/sec: 410999626.38
[root@ceph-iscsi1 mnt]# rbd bench --image testec --io-size 64K --io-type write --io-total 10G
bench type write io_size 65536 io_threads 16 bytes 10737418240 pattern sequential
SEC OPS OPS/SEC BYTES/SEC
1 7392 7415.38 485974266.41
2 14464 7243.59 474715656.29
3 22000 7341.08 481104853.50
4 29408 7352.29 481839517.16
5 37296 7459.38 488857889.75
6 44864 7494.36 491150574.57
7 52848 7676.76 503104281.98
8 60784 7756.76 508347136.11
9 68608 7835.26 513491609.52
10 76784 7902.30 517885290.67
11 84544 7935.96 520091129.45
12 92432 7916.76 518832844.57
13 100064 7855.96 514848275.43
14 107040 7692.52 504136734.09
15 114320 7499.66 491497933.56
16 121744 7436.99 487390477.85
17 129664 7438.92 487517345.01
18 136704 7326.50 480149408.39
19 144960 7587.00 497221460.09
20 153264 7796.56 510955233.33
21 160832 7814.44 512126854.90
elapsed: 21 ops: 163840 ops/sec: 7659.97 bytes/sec: 502004079.43
On Fri, Oct 25, 2019 at 11:54 AM Mike Christie <mchristi@xxxxxxxxxx> wrote:
On 10/24/2019 11:47 PM, Ryan wrote:
I'm using CentOS 7.7.1908 with kernel 3.10.0-1062.1.2.el7.x86_64. The
workload was a VMware Storage Motion from a local SSD backed datastore
Ignore my comments. I thought you were just doing fio like tests in the vm.
to the ceph backed datastore. Performance was measured using dstat on
the iscsi gateway for network traffic and ceph status as this cluster is
basically idle. I changed max_data_area_mb to 256 and cmdsn_depth to
128. This appears to have given a slight improvement of maybe 10MB/s.
Moving VM to the ceph backed datastore
io:
client: 124 KiB/s rd, 76 MiB/s wr, 95 op/s rd, 1.26k op/s wr
Moving VM off the ceph backed datastore
io:
client: 344 MiB/s rd, 625 KiB/s wr, 5.54k op/s rd, 62 op/s wr
If you run esxtop while running your test what do you see for the number
of commands in the iscsi LUN's queue?
I'm going to test bonnie++ with an rbd volume mounted directly on the
To try and isolate if its the iscsi or rbd, you need to run fio with the
librbd io engine. We know krbd is going to be the fastest. ceph-iscsi
uses librbd so it is a better baseline. If you are not familiar with fio
you can just do something like:
fio --group_reporting --ioengine=rbd --direct=1 --name=librbdtest
--numjobs=32 --bs=512k --iodepth=128 --size=10G --rw=write
--rbd=name_of_your_image -pool=name_of_pool
iscsi gateway. Also will test bonnie++ inside a VM on a ceph backed
datastore.
On Thu, Oct 24, 2019 at 7:15 PM Mike Christie <mchristi@xxxxxxxxxx
<mailto:mchristi@xxxxxxxxxx>> wrote:
On 10/24/2019 12:22 PM, Ryan wrote:
> I'm in the process of testing the iscsi target feature of ceph. The
> cluster is running ceph 14.2.4 and ceph-iscsi 3.3. It consists of 5
What kernel are you using?
> hosts with 12 SSD OSDs per host. Some basic testing moving VMs to
a ceph
> backed datastore is only showing 60MB/s transfers. However moving
these
> back off the datastore is fast at 200-300MB/s.
What is the workload and what are you using to measure the throughput?
If you are using fio, what arguments are you using? And, could you
change the ioengine to rbd and re-run the test from the target system so
we can check if rbd is slow or iscsi?
For small IOs, 60 is about right.
For 128-512K IOs you should be able to get around 300 MB/s for writes
and 600 for reads.
1. Increase max_data_area_mb. This is a kernel buffer lio/tcmu uses to
pass data between the kernel and tcmu-runner. The default is only 8MB.
In gwcli cd to your disk and do:
# reconfigure max_data_area_mb %N
where N is between 8 and 2048 MBs.
2. The Linux kernel target only allows 64 commands per iscsi session by
default. We increase that to 128, but you can increase this to 512.
In gwcli cd to the target dir and do
reconfigure cmdsn_depth 512
3. I think ceph-iscsi and lio work better with higher queue depths so if
you are using fio you want higher numjobs and/or iodepths.
>
> What should I be looking at to track down the write performance issue?
> In comparison with the Nimble Storage arrays I can see 200-300MB/s in
> both directions.
>
> Thanks,
> Ryan
>
>
> _______________________________________________
> ceph-users mailing list -- ceph-users@xxxxxxx
<mailto:ceph-users@xxxxxxx>
> To unsubscribe send an email to ceph-users-leave@xxxxxxx
<mailto:ceph-users-leave@xxxxxxx>
>
_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx
To unsubscribe send an email to ceph-users-leave@xxxxxxx
_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx
To unsubscribe send an email to ceph-users-leave@xxxxxxx
_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx
To unsubscribe send an email to ceph-users-leave@xxxxxxx
