[Lustre-discuss] Lustre and system cpu usage

[Claudio Baeza Retamal]

Dears folks,

last month, I have configured lustre 1.8.5 over infiniband, before, I
was using Gluster 3.1.2, performance was ok but reliability was wrong,
when 40 or more applications requested at the same time for open a
file, gluster servers bounced randomly the active connections from
clients. Lustre has not this problem, but I can see others issues, for
example, namd appears with "system cpu" around of 30%, hpl benchmark
appears between 70%-80% of "system cpu", is too much high, with
gluster, the system cpu was never exceeded 5%. I think, this is
explained due gluster uses fuse and run in user space, but I am do not
sure. I have some doubt:

¿why Lustre uses ipoib? Before, with gluster I do not use ipoib, I am
thinking that ipoib module produces bad performance in infiniband and
disturbs the infiniband native module.
It is posible to configure lustre to transport metada over ethernet and
data over infiniband?

For namd and hpl benchmark, is it normal to have system cpu to be so high?

My configuration is the following:

- Qlogic 12800-180 switch, 7 leaf (24 ports per leaf) and 2 spines (All
ports have QDR, 40 Gbps)
- 66 HCA mellanox connectX, two ports, QDR 40 Gbps (compute nodes)
- 1 metadata server, 96 GB RAM DDR3 optimized for performance, two Xeon
5570, SAS 15K RPM hard disk in Raid 1, HCA mellanox connectX with two ports
- 4 OSS with 1 OST of 2 TB in RAID 5 each one (8 TB in total). The all
OSS have a Mellanox ConnectX with two ports

I will appreciate any help or tips.

Regards

claudio

--
Claudio Baeza Retamal
CTO
National Laboratory for High Performance Computing (NLHPC)
Center for Mathematical Modeling (CMM)
School of Engineering and Sciences
Universidad de Chile

_______________________________________________
Lustre-discuss mailing list
Lustre-...@lists.lustre.org
http://lists.lustre.org/mailman/listinfo/lustre-discuss

[Andreas Dilger]

On 2011-01-14, at 3:57 PM, Claudio Baeza Retamal wrote:
> last month, I have configured lustre 1.8.5 over infiniband, before, I
> was using Gluster 3.1.2, performance was ok but reliability was wrong,
> when 40 or more applications requested at the same time for open a
> file, gluster servers bounced randomly the active connections from
> clients. Lustre has not this problem, but I can see others issues, for
> example, namd appears with "system cpu" around of 30%, hpl benchmark
> appears between 70%-80% of "system cpu", is too much high, with
> gluster, the system cpu was never exceeded 5%. I think, this is
> explained due gluster uses fuse and run in user space, but I am do not
> sure.

If Gluster is using FUSE, then all of the CPU usage would appear in "user" and not "system". That doesn't mean that the CPU usage is gone, just accounted in a different place.

> I have some doubt:
>
> ¿why Lustre uses ipoib? Before, with gluster I do not use ipoib, I am
> thinking that ipoib module produces bad performance in infiniband and
> disturbs the infiniband native module.

If you are using IPoIB for data then your LNET is configured incorrectly. IPoIB is only needed for IB hostname resolution, and all LNET traffic can use native IB with very low CPU overhead. Your /etc/modprobe.conf and mount lines should be using {addr}@o2ib0 instead of {addr} or {addr}@tcp0.

> It is posible to configure lustre to transport metada over ethernet and
> data over infiniband?

Yes, this should be possible, but putting the metadata on IB is much lower latency and higher performance so you should really try to use IB for both.

> For namd and hpl benchmark, is it normal to have system cpu to be so high?
>
> My configuration is the following:
>
> - Qlogic 12800-180 switch, 7 leaf (24 ports per leaf) and 2 spines (All
> ports have QDR, 40 Gbps)
> - 66 HCA mellanox connectX, two ports, QDR 40 Gbps (compute nodes)
> - 1 metadata server, 96 GB RAM DDR3 optimized for performance, two Xeon
> 5570, SAS 15K RPM hard disk in Raid 1, HCA mellanox connectX with two ports
> - 4 OSS with 1 OST of 2 TB in RAID 5 each one (8 TB in total). The all
> OSS have a Mellanox ConnectX with two ports

If you have IB on the MDS then you should definitely use {addr}@o2ib0 for both OSS and MDS nodes. That will give you much better metadata performance.

Cheers, Andreas
--
Andreas Dilger
Principal Engineer
Whamcloud, Inc.

[Peter Kjellström]

On Saturday, January 15, 2011 12:57:18 am Claudio Baeza Retamal wrote:
> Dears folks,
>
> last month, I have configured lustre 1.8.5 over infiniband, before, I
> was using Gluster 3.1.2, performance was ok but reliability was wrong,
> when 40 or more applications requested at the same time for open a
> file, gluster servers bounced randomly the active connections from
> clients. Lustre has not this problem, but I can see others issues, for
> example, namd appears with "system cpu" around of 30%, hpl benchmark
> appears between 70%-80% of "system cpu",

This is certainly an indication of something being very wrong (not necessarily
lustre). hpl does very minimal I/O and only a completely broken filesystem
would cause 70-80% "system cpu".

/Peter

[Claudio Baeza Retamal]

Hi,

El 14-03-2011 22:05, Andreas Dilger escribió:
> On 2011-01-14, at 3:57 PM, Claudio Baeza Retamal wrote:
>> last month, I have configured lustre 1.8.5 over infiniband, before, I
>> was using Gluster 3.1.2, performance was ok but reliability was wrong,
>> when 40 or more applications requested at the same time for open a
>> file, gluster servers bounced randomly the active connections from
>> clients. Lustre has not this problem, but I can see others issues, for
>> example, namd appears with "system cpu" around of 30%, hpl benchmark
>> appears between 70%-80% of "system cpu", is too much high, with
>> gluster, the system cpu was never exceeded 5%. I think, this is
>> explained due gluster uses fuse and run in user space, but I am do not
>> sure.
> If Gluster is using FUSE, then all of the CPU usage would appear in "user" and not "system". That doesn't mean that the CPU usage is gone, just accounted in a different place.
>
>
>> I have some doubt:
>>
>> ¿why Lustre uses ipoib? Before, with gluster I do not use ipoib, I am
>> thinking that ipoib module produces bad performance in infiniband and
>> disturbs the infiniband native module.
> If you are using IPoIB for data then your LNET is configured incorrectly. IPoIB is only needed for IB hostname resolution, and all LNET traffic can use native IB with very low CPU overhead. Your /etc/modprobe.conf and mount lines should be using {addr}@o2ib0 instead of {addr} or {addr}@tcp0.
>

For first two weeks, I was using "options lnet networks="o2ib(ib0)",
now, I am using "options lnet networks="o2ib(ib0),tcp0(eth0)" because I
have a node without HCA card, in both case, the system cpu usage is the
same, the compute node without infiniband is used to run matlab only.

In the hpl benchmark case, my doubt is, why has a high system cpu
usage? Is posible that LustreFS disturbs mlx4 infiniband driver and
causes problems with MPI? hpl benchmark mainly does I/O for transport
data over MPI, with glusterFS system cpu was around 5%, instead, since
Lustre was configured system cpu is 70%-80% and we use o2ib(ib0) for
LNET in modprobe.conf .
I have tried several options, following instruction from mellanox, in
compute nodes I disable irqbalance and run smp_affinity script, but
system cpu still so higher.
Are there any tools to study lustre performance?

>> It is posible to configure lustre to transport metada over ethernet and
>> data over infiniband?
> Yes, this should be possible, but putting the metadata on IB is much lower latency and higher performance so you should really try to use IB for both.
>
>> For namd and hpl benchmark, is it normal to have system cpu to be so high?
>>
>> My configuration is the following:
>>
>> - Qlogic 12800-180 switch, 7 leaf (24 ports per leaf) and 2 spines (All
>> ports have QDR, 40 Gbps)
>> - 66 HCA mellanox connectX, two ports, QDR 40 Gbps (compute nodes)
>> - 1 metadata server, 96 GB RAM DDR3 optimized for performance, two Xeon
>> 5570, SAS 15K RPM hard disk in Raid 1, HCA mellanox connectX with two ports
>> - 4 OSS with 1 OST of 2 TB in RAID 5 each one (8 TB in total). The all
>> OSS have a Mellanox ConnectX with two ports
> If you have IB on the MDS then you should definitely use {addr}@o2ib0 for both OSS and MDS nodes. That will give you much better metadata performance.
>
> Cheers, Andreas
> --
> Andreas Dilger
> Principal Engineer
> Whamcloud, Inc.
>
>
>
>
>

regards

claudio

[Andreas Dilger]

On 2011-01-15, at 4:18 AM, Claudio Baeza Retamal wrote:
> El 14-03-2011 22:05, Andreas Dilger escribió:
>> On 2011-01-14, at 3:57 PM, Claudio Baeza Retamal wrote:
>>> last month, I have configured lustre 1.8.5 over infiniband, before, I
>>> was using Gluster 3.1.2, performance was ok but reliability was wrong,
>>> when 40 or more applications requested at the same time for open a
>>> file, gluster servers bounced randomly the active connections from
>>> clients. Lustre has not this problem, but I can see others issues, for
>>> example, namd appears with "system cpu" around of 30%, hpl benchmark
>>> appears between 70%-80% of "system cpu", is too much high, with
>>> gluster, the system cpu was never exceeded 5%. I think, this is
>>> explained due gluster uses fuse and run in user space, but I am do not
>>> sure.

>>> I have some doubt:
>>>
>>> ¿why Lustre uses ipoib? Before, with gluster I do not use ipoib, I am
>>> thinking that ipoib module produces bad performance in infiniband and
>>> disturbs the infiniband native module.
>>
>> If you are using IPoIB for data then your LNET is configured incorrectly. IPoIB is only needed for IB hostname resolution, and all LNET traffic can use native IB with very low CPU overhead. Your /etc/modprobe.conf and mount lines should be using {addr}@o2ib0 instead of {addr} or {addr}@tcp0.
>
> For first two weeks, I was using "options lnet networks="o2ib(ib0)", now, I am using "options lnet networks="o2ib(ib0),tcp0(eth0)" because I have a node without HCA card, in both case, the system cpu usage is the same, the compute node without infiniband is used to run matlab only.
>
> In the hpl benchmark case, my doubt is, why has a high system cpu usage? Is posible that LustreFS disturbs mlx4 infiniband driver and causes problems with MPI? hpl benchmark mainly does I/O for transport data over MPI, with glusterFS system cpu was around 5%, instead, since Lustre was configured system cpu is 70%-80% and we use o2ib(ib0) for LNET in modprobe.conf.

Have you tried disabling the Lustre kernel debug logs (lctl set_param debug=0) and/or disabling the network data checksums (lctl set_param osc.*.checksums=0)?

Note that there is also CPU overhead in the kernel from copying data from userspace to the kernel that is unavoidable for any filesystem, unless O_DIRECT is used (which causes synchronous IO and has IO alignment restrictions).

[Wojciech Turek]

Hi Claudio,

When you say that during Linpack you see high system cpu usage, do you mean the cpu usage on the clients or the servers?
Can you run for example top command and see which processes take the most of the CPU time?

Cheers

Wojciech