IOTrap

Client side iotrap infrastructure to intercept and enhance IO performance

Overview

The iotrap initiative aims at trapping the client side IO requests (currently supports io_submit()) and sends them over multiple mounts (connections) over NFS. The increased connections translates to an increase in network threads available for the client. This allows us to achieve better throughput compared to a single connection/thread

Performance improvement seen with READ over 10g network: NFSv3 vanilla: 718mb/s NFSv3 with iotrap: 1100mb/s NFSv4 vanilla: 523mb/s NFSv4 with iotrap: 1094mb/s

=> Almost maxing out the network bandwidth with single 'fsperf' instance

Setup

Download iotrap.so (say under: /root/iotrap.so)

Mount the same volume (nfs_perf_rhelx01) from different data LIFs (preferably 4 LIFs), example (my test setup):

172.17.90.32:/nfs_perf_rhelx01 on /root/hana/mnt/v32
172.17.90.33:/nfs_perf_rhelx01 on /root/hana/mnt/v33
172.17.90.34:/nfs_perf_rhelx01 on /root/hana/mnt/v34
172.17.90.35:/nfs_perf_rhelx01 on /root/hana/mnt/v35

All the data LIFs can be on the same port on the filer (so that they use the same network interface), example:

vserver	data LIF	home-port	address
vs0	nfs-f8060x05_data1	e4a	172.17.90.32
vs0	nfs-f8060x05_data2	e4a	172.17.90.33
vs0	nfs-f8060x05_data3	e4a	172.17.90.34
vs0	nfs-f8060x05_data4	e4a	172.17.90.35

On the client side, in the shell you intend to run the tests: Set environment variable IOTRAP_MOUNT_POOLS (some_name:comma,separated,mount,paths):

$ export IOTRAP_MOUNT_POOLS="e4a:/root/hana/mnt/v32/,/root/hana/mnt/v33/,/root/hana/mnt/v34/,/root/hana/mnt/v35/ e4b:/root/hana/mnt/v42/,/root/hana/mnt/v43/,/root/hana/mnt/v44/,/root/hana/mnt/v45/ e3a:/root/hana/mnt/v52/,/root/hana/mnt/v53/,/root/hana/mnt/v54/,/root/hana/mnt/v55/ e3b:/root/hana/mnt/v62/,/root/hana/mnt/v63/,/root/hana/mnt/v64/,/root/hana/mnt/v65/"

Intercepted funcations can be traced by setting IOTRAP_TRACE environment variable. To stop tracking, unset the environment variable and run the app

Running the SAP Hana 'fsperf' test under GNU/Linux

Without the iotrap improvements

$ echo 3 > /proc/sys/vm/drop_caches
$ ./fsperf -t read -m throughput -f 2g -b 64k -o short --noinit /root/hana/mnt/v32/

With iotrap based improvements

$ echo 3 > /proc/sys/vm/drop_caches
$ LD_PRELOAD=/root/iotrap.so ./fsperf -t read -m throughput -f 2g -b 64k -o short --noinit /root/hana/mnt/v32/

jbd / iotrap