The file removal performance is very low by mdtest


Groot
 

I install the DAOS and use the mdtest to do some test with -a DFS (one server and one client with 20 cores, the server use 256G DCPM without nvme ssd). But I find the file removal performance is very low and the result shown as follow and I want to know if it is right? Or how to do the performance turning?


The config file of server is :

access_points: ['server-1']

port: 10001

transport_config:

    allow_insecure: false

    client_cert_dir: /etc/daos/certs/clients

    ca_cert: /etc/daos/certs/daosCA.crt

    cert: /etc/daos/certs/server.crt

    key: /etc/daos/certs/server.key

provider: ofi+verbs;ofi_rxm

control_log_mask: INFO

control_log_file: /tmp/daos_server.log

helper_log_file: /tmp/daos_admin.log

telemetry_port: 9191

 

engines:

-

    targets: 20

    nr_xs_helpers: 20

    fabric_iface: ib0

    fabric_iface_port: 31316

    log_mask: ERR

    log_file: /tmp/daos_engine_0.log

    scm_mount: /mnt/daos0

    scm_class: dcpm

    scm_list: [/dev/pmem0]

Thanks a lot
Groot


Chaarawi, Mohamad
 

Hi,

 

What is you mdtest command that you used to run?

Typically low remove performance might indicate that you are using SX object class for files (the default),, which is fine if your files are going to be large. However that’s not the case with mdtest.

So I would suggest to re-run the same command, but add:

--dfs.oclass S1 --dfs.dir_oclass SX

This tells DFS to use:

  • single stripe/shard for files (best for mdtest where files are small or empty)
  • full striping for directories (also best for mdtest where directories usually have a lot of files).

 

Thanks,

Mohamad

 

From: daos@daos.groups.io <daos@daos.groups.io> on behalf of Groot <kukougu@...>
Date: Tuesday, April 12, 2022 at 10:15 PM
To: daos@daos.groups.io <daos@daos.groups.io>
Subject: [daos] The file removal performance is very low by mdtest

I install the DAOS and use the mdtest to do some test with -a DFS (one server and one client with 20 cores, the server use 256G DCPM without nvme ssd). But I find the file removal performance is very low and the result shown as follow and I want to know if it is right? Or how to do the performance turning?


The config file of server is :

access_points: ['server-1']

port: 10001

transport_config:

    allow_insecure: false

    client_cert_dir: /etc/daos/certs/clients

    ca_cert: /etc/daos/certs/daosCA.crt

    cert: /etc/daos/certs/server.crt

    key: /etc/daos/certs/server.key

provider: ofi+verbs;ofi_rxm

control_log_mask: INFO

control_log_file: /tmp/daos_server.log

helper_log_file: /tmp/daos_admin.log

telemetry_port: 9191

 

engines:

-

    targets: 20

    nr_xs_helpers: 20

    fabric_iface: ib0

    fabric_iface_port: 31316

    log_mask: ERR

    log_file: /tmp/daos_engine_0.log

    scm_mount: /mnt/daos0

    scm_class: dcpm

    scm_list: [/dev/pmem0]

Thanks a lot
Groot


Groot
 

Thanks, I want to know where I can get the description of --dfs.oclass and --dfs.dir_oclass?Do they contain other parameters except S1 and SX?


Wang, Shilong
 

Try to explicitly set the object class for files and directories –

the defaults are not providing the best performance for mdtest-like workloads:

--dfs.oclass=S1 --dfs.dir_oclass=SX

 

From: daos@daos.groups.io <daos@daos.groups.io> On Behalf Of Groot
Sent: Wednesday, April 13, 2022 11:15 AM
To: daos@daos.groups.io
Subject: [daos] The file removal performance is very low by mdtest

 

I install the DAOS and use the mdtest to do some test with -a DFS (one server and one client with 20 cores, the server use 256G DCPM without nvme ssd). But I find the file removal performance is very low and the result shown as follow and I want to know if it is right? Or how to do the performance turning?


The config file of server is :

access_points: ['server-1']

port: 10001

transport_config:

    allow_insecure: false

    client_cert_dir: /etc/daos/certs/clients

    ca_cert: /etc/daos/certs/daosCA.crt

    cert: /etc/daos/certs/server.crt

    key: /etc/daos/certs/server.key

provider: ofi+verbs;ofi_rxm

control_log_mask: INFO

control_log_file: /tmp/daos_server.log

helper_log_file: /tmp/daos_admin.log

telemetry_port: 9191

 

engines:

-

    targets: 20

    nr_xs_helpers: 20

    fabric_iface: ib0

    fabric_iface_port: 31316

    log_mask: ERR

    log_file: /tmp/daos_engine_0.log

    scm_mount: /mnt/daos0

    scm_class: dcpm

    scm_list: [/dev/pmem0]

Thanks a lot
Groot