Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
Hi guys,
I am testing the performance of DRBD transport added in DRBD9. The test
result seems problematic because
it's even worse than ISCSI transport.
Here's the test result of 4k random write on SSD disk with fio.
Raw disk DRBD8 Local DRBD9 Local DRBD Transport
ISCSI Transport with DRBD9
IOPS 79756 55549 28624 16343
33435
Latency(us) 199.65 286.36 138.43
977.73 463.08
Please note the "DRBD8 Local" is tested with a tuned configuration while
all DRBD9 tests are tested with the configurations generated by drbdmanage.
It's assumed that DRBD transport should give better performance than ISCSI
transport. I don't know why it's even worse. Any suggestion to tune the
DRBD transport? Thanks!
The following is my test setup and configurations:
Disk: SSD Samsung SSD 845DC EVO 480GB
NIC: 10Gb
OS Kernel: 3.10.0-327.10.1.el7
DRBD9 kernel module: 9.0.1-1 (GIT-hash:
f57acfc22d29a95697e683fb6bbacd9a1ad4348e)
drbdmanage-0.92
drbd-utils-8.9.6
DRBD8 tuned configurations are copied from
http://www.linbit.com/en/resources/technical-publications/101-performance/545-intel-sata-ssd-testing-with-drbd-2
DRBD9 volume configuration (The nodes drbd1 and drbd2 provide the storage
of volume vol2 and the drbd3 is the diskless client)
[root at drbd1 ~]# drbdsetup show vol2
resource vol2 {
_this_host {
node-id 0;
volume 0 {
device minor 102;
disk "/dev/drbdpool/vol2_00";
meta-disk internal;
disk {
size 39062500s; # bytes
disk-flushes no;
md-flushes no;
al-extents 6433;
al-updates no;
read-balancing least-pending;
}
}
}
connection {
_peer_node_id 1;
path {
_this_host ipv4 192.168.253.131:7001;
_remote_host ipv4 192.168.253.132:7001;
}
net {
max-epoch-size 20000;
sndbuf-size 524288; # bytes
cram-hmac-alg "sha1";
shared-secret "q9Ku9G/Z/fhG1b3aemcD";
verify-alg "sha1";
max-buffers 81920;
_name "drbd2";
}
volume 0 {
disk {
resync-rate 409600k; # bytes/second
c-plan-ahead 10; # 1/10 seconds
c-fill-target 88s; # bytes
c-max-rate 614400k; # bytes/second
c-min-rate 10240k; # bytes/second
}
}
}
}
[root at drbd3 ~]# drbdsetup show vol2
resource vol2 {
_this_host {
node-id 2;
volume 0 {
device minor 102;
}
}
connection {
_peer_node_id 0;
path {
_this_host ipv4 192.168.253.133:7001;
_remote_host ipv4 192.168.253.131:7001;
}
net {
cram-hmac-alg "sha1";
shared-secret "q9Ku9G/Z/fhG1b3aemcD";
_name "drbd1";
}
}
connection {
_peer_node_id 1;
path {
_this_host ipv4 192.168.253.133:7001;
_remote_host ipv4 192.168.253.132:7001;
}
net {
cram-hmac-alg "sha1";
shared-secret "q9Ku9G/Z/fhG1b3aemcD";
_name "drbd2";
}
}
}
The iscsi target is exported on drbd2 via LIO. I switched the primary from
drbd3 to drbd2 when testing the iscsi transport.
FIO test configuration
[global]
bs=4k
ioengine=libaio
iodepth=16
size=10g
direct=1
runtime=300
directory=/mnt/ssd1
filename=ssd.test.file
[rand-write]
rw=randwrite
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.linbit.com/pipermail/drbd-user/attachments/20160225/bbed5539/attachment.htm>