Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
I asked for the same thing (without SSD) a few weeks ago.
Someone answered me that these preformances are perfectly normal is dual master
configuration.
Seems to be due to the network latency (first server I/O + network latency +
second server I/O + network latency (ACK))
I finally decided that DRBD is unusable in dual primary setup because the
performance drop.
Julien
Le 29/10/2010 18:09, wang xuchen a écrit :
> Hi all,
>
> I have encountered a DRBD write performance bottleneck issue.
>
> According to DRBD specification "DRBD then reduces that throughput maximum by
> its additional throughput overhead, which can be expected to beless than 3 percent."
>
> My current test environment is:
>
> (1) Hard-drive: 300G SSD with 8 partitions on it, each of which has a DRBD
> device create on top it. I use dd utility to test its performance: 97 MB/s with
> 4k block size.
>
>
> (2) netowork: dedicated 10G ethernet card for data replication:
> ethtool eth2
> Settings for eth2:
> ...
> Speed: 10000Mb/s
> ...
>
> (3) DRBD configuration: (Here is one of them).
>
> on Server1 {
> device /dev/drbd3 minor 3;
> disk /dev/fioa3;
> address ipv4 192.168.202.107:7793 <http://192.168.202.107:7793>;
> meta-disk internal;
> }
> on NSS_108 {
> device /dev/drbd3 minor 3;
> disk /dev/fioa3;
> address ipv4 192.168.202.108:7793 <http://192.168.202.108:7793>;
> meta-disk internal;
> }
> net {
> allow-two-primaries;
> after-sb-0pri discard-zero-changes;
> after-sb-1pri consensus;
> after-sb-2pri call-pri-lost-after-sb;
> rr-conflict disconnect;
> max-buffers 4000;
> max-epoch-size 16000;
> unplug-watermark 4000;
> sndbuf-size 2M;
> data-integrity-alg crc32c;
> }
> syncer {
> rate 300M;
> csums-alg md5;
> verify-alg crc32c;
> al-extents 3800;
> cpu-mask 2;
> }
> }
>
> (4) Test result:
>
> I have a simple script which use multiple instance of dd to their corresponding
> DRBD device
>
> dd if=/dev/zero of=/dev/drbd1 bs=4k count=10000 oflag=direct &
> ....
>
> For one device, I got roughly 8M/s. As the test goes, I increase the number of
> device to see if it helps the performance. Unfortunately, as the number of
> device grows, performance seems to be distributed on each of the device with the
> total add up to 10M/s.
>
> Can somebody give me a hint on what was going wrong?
>
> Many Thanks.
> Ben
>
>
> Commit yourself to constant self-improvement