Skip to contentSkip to conversation

SRP subnet timeout

DISCUSSION Karandeep Chahal 2 participants 7 messages Sep 26, 2019 – Sep 27, 2019 InfiniBand

lore
Share & tools
lore
#0 lore raw

Hi All,

I am wondering why SRP REQ CM timeout is "subnet_timeout + 2". What was the reason behind adding 2 to the subnet timeout? I think it was added by Bart in commit 4c532d6ce14bb3fb4e1cb2d29fafdd7d6bded51c, but the commit message does not explain why the 2 is necessary. Would anyone happen to know?

My understanding is that adding 2 causes the CM timeout value to become 4 times the subnet_timeout value. Hence, even if you have a reasonable subnet_timeout the CM REQ timeout becomes too high.

Thank
-Karan

#1 lore raw
Quoted 8 lines · Karandeep Chahal

I am wondering why SRP REQ CM timeout is "subnet_timeout + 2". What was the reason behind adding 2 to the subnet timeout? I think it was added by Bart in commit 4c532d6ce14bb3fb4e1cb2d29fafdd7d6bded51c, but the commit message does not explain why the 2 is necessary. Would anyone happen to know?

My understanding is that adding 2 causes the CM timeout value to become 4 times the subnet_timeout value. Hence, even if you have a reasonable subnet_timeout the CM REQ timeout becomes too high.

Hi Karan,

I think the description of that commit explains why that change has been made:

IB/srp: Make CM timeout dependent on subnet timeout

For small networks it is safe to reduce the subnet timeout from its default value (18 for opensm) to 16. Make the SRP CM timeout dependent on the subnet timeout such that decreasing the subnet timeout also causes SRP failover and failback to occur faster.

Bart.

#2 lore raw

Hi Bart,

I understand that you have made "SRP CM timeout dependent on subnet timeout", but my question is, why did you add "2" to the subnet timeout? How did you come up with this number?

Forgive me but it is not obvious to me after reading the commit message.

Thanks
-Karan

Quoted 22 lines · Bart Van Assche
Quoted 8 lines · Karandeep Chahal

I am wondering why SRP REQ CM timeout is "subnet_timeout + 2". What was the reason behind adding 2 to the subnet timeout? I think it was added by Bart in commit 4c532d6ce14bb3fb4e1cb2d29fafdd7d6bded51c, but the commit message does not explain why the 2 is necessary. Would anyone happen to know?

My understanding is that adding 2 causes the CM timeout value to become 4 times the subnet_timeout value. Hence, even if you have a reasonable subnet_timeout the CM REQ timeout becomes too high.

Hi Karan,

I think the description of that commit explains why that change has been made:

    IB/srp: Make CM timeout dependent on subnet timeout

    For small networks it is safe to reduce the subnet timeout from its default value (18 for opensm) to 16. Make the SRP CM timeout dependent on the subnet timeout such that decreasing the subnet timeout also causes SRP failover and failback to occur faster.

Bart.

#3 lore raw

A: Because it messes up the order in which people normally read text.
Q: Why is top-posting such a bad thing?
A: Top-posting.
Q: What is the most annoying thing in e-mail?

Quoted 5 lines · Karandeep Chahal

I understand that you have made "SRP CM timeout dependent on subnet timeout", but my question is, why did you add "2" to the subnet timeout? How did you come up with this number?

Forgive me but it is not obvious to me after reading the commit message.

Hi Karandeep,

I haven't done any kind of deep research to come up with that value. My goal was to preserve the CM timeout used by the SRP initiator for the default opensm subnet timeout value (18). There may be better choices for the SRP CM timeout.

Bart.

#4 lore raw
Quoted 11 lines · Bart Van Assche

A: Because it messes up the order in which people normally read text.
Q: Why is top-posting such a bad thing?
A: Top-posting.
Q: What is the most annoying thing in e-mail?

Quoted 4 lines · Karandeep Chahal

I understand that you have made "SRP CM timeout dependent on subnet timeout", but my question is, why did you add "2" to the subnet timeout? How did you come up with this number?

Quoted 7 lines
Quoted 1 line

Forgive me but it is not obvious to me after reading the commit message.

Hi Karandeep,

I haven't done any kind of deep research to come up with that value. My goal was to preserve the CM timeout used by the SRP initiator for the default opensm subnet timeout value (18). There may be better choices for the SRP CM timeout.

Bart.

Hi Bart,

My apologies for top-posting.

The problem with adding 2 to SRP CM timeout is that it makes the SRP CM timeout value far off the subnet timeout value.
Now you can either have a reasonable subnet timeout value and an unreasonable SRP CM timeout value, or an unreasonable subnet timeout and a reasonable SRP CM timeout.

-Karan

#5 lore raw
Quoted 5 lines · Karandeep Chahal

The problem with adding 2 to SRP CM timeout is that it makes the SRP CM timeout value far off the subnet timeout value.
Now you can either have a reasonable subnet timeout value and an unreasonable SRP CM timeout value, or an unreasonable subnet timeout and a reasonable SRP CM timeout.

Hi Karan,

Do you have a proposal for how the SRP initiator driver should be modified (patch)? After a patch is available the next step is to write a text that describes why that change will work for all users (patch description).

Thanks,

Bart.

#6 lore raw
Quoted 8 lines · Bart Van Assche
Quoted 5 lines · Karandeep Chahal

The problem with adding 2 to SRP CM timeout is that it makes the SRP CM timeout value far off the subnet timeout value.
Now you can either have a reasonable subnet timeout value and an unreasonable SRP CM timeout value, or an unreasonable subnet timeout and a reasonable SRP CM timeout.

Hi Karan,

Quoted 8 lines

Do you have a proposal for how the SRP initiator driver should be modified (patch)? After a patch is available the next step is to write a text that describes why that change will work for all users (patch description).

Thanks,

Bart.

Hi Bart,

I will come up with a patch/proposal sometime next week. Please stay tuned.

Thanks
-Karan