kerrighed 3.0 testing using wired and wireless nodes

[lokesh mandvekar]

We ran a sample R program using multicores over 3 different setups
(each using the same 3 identical machines) and got the corresponding
execution times:

1. 3 nodes with wired connection - execution time: 220 seconds
2. only 2 nodes wired - execution time: 292 seconds
3. 2 nodes wired and the 3rd node wireless (802.11b) - execution time:
493 seconds

I do recollect being told that the min bandwidth requirement for
Kerrighed is 100Mbps. But considering cases 2 and 3 above, is the
increase in execution time by adding a wireless node expected?

It would be great if I could be directed to any specific areas in the
Kerrighed source code which are influenced by bandwidth. Also, if
there have been prior tests performed over wireless with favorable
results, I'd be happy to know about those.

Please do let me know if anything's unclear or if any further details required.

Thanks,

[Louis Rilling]

Hi,

Post by lokesh mandvekar
We ran a sample R program using multicores over 3 different setups
(each using the same 3 identical machines) and got the corresponding
1. 3 nodes with wired connection - execution time: 220 seconds
2. only 2 nodes wired - execution time: 292 seconds
493 seconds

First of all, thank you for your feedback.

Post by lokesh mandvekar
I do recollect being told that the min bandwidth requirement for
Kerrighed is 100Mbps. But considering cases 2 and 3 above, is the
increase in execution time by adding a wireless node expected?

Although bandwidth is important, Kerrighed is very sensitive to latency. My
guess is that the latency of wireless is so high that migrating/remote forking
small(?) processes over a wireless link decreases the overall performance.

Post by lokesh mandvekar
It would be great if I could be directed to any specific areas in the
Kerrighed source code which are influenced by bandwidth. Also, if
there have been prior tests performed over wireless with favorable
results, I'd be happy to know about those.

I don't know about tests over wireless.

As fas as bandwidth is concerned, the most relevant area is memory management
(that is in kerrighed/mm/ and the underlying KDDM layer in kddm/): on a
migration / remote fork, 4k pages of the process address space are transeferred
on demand (that is when accessed from the target node) from the source node to
the target node.

However, I think that the issue with wireless is more latency than bandwidth.
I would say that all Kerrighed kernel-to-kernel communications are
latency-sensitive.

Post by lokesh mandvekar
Please do let me know if anything's unclear or if any further details required.

It looks pretty clear. Showing the bandwidth/latency of the different network
fabrics, as well as measuring migration/remote fork time + the memory size of
your processes would complete the picture.

Thanks,

Louis

[lokesh mandvekar]

Post by Louis Rilling
Hi,

Post by lokesh mandvekar
We ran a sample R program using multicores over 3 different setups
(each using the same 3 identical machines) and got the corresponding
1. 3 nodes with wired connection - execution time: 220 seconds
2. only 2 nodes wired - execution time: 292 seconds
493 seconds

First of all, thank you for your feedback.

Post by lokesh mandvekar
I do recollect being told that the min bandwidth requirement for
Kerrighed is 100Mbps. But considering cases 2 and 3 above, is the
increase in execution time by adding a wireless node expected?

Although bandwidth is important, Kerrighed is very sensitive to latency. My
guess is that the latency of wireless is so high that migrating/remote forking
small(?) processes over a wireless link decreases the overall performance.

Post by lokesh mandvekar
It would be great if I could be directed to any specific areas in the
Kerrighed source code which are influenced by bandwidth. Also, if
there have been prior tests performed over wireless with favorable
results, I'd be happy to know about those.

I don't know about tests over wireless.
As fas as bandwidth is concerned, the most relevant area is memory management
(that is in kerrighed/mm/ and the underlying KDDM layer in kddm/): on a
migration / remote fork, 4k pages of the process address space are transeferred
on demand (that is when accessed from the target node) from the source node to
the target node.
However, I think that the issue with wireless is more latency than bandwidth.
I would say that all Kerrighed kernel-to-kernel communications are
latency-sensitive.

Post by lokesh mandvekar
Please do let me know if anything's unclear or if any further details required.

It looks pretty clear. Showing the bandwidth/latency of the different network
fabrics, as well as measuring migration/remote fork time + the memory size of
your processes would complete the picture.

Thanks a lot for the reply. I'll try looking into it and get the
required data asap.

Post by Louis Rilling
Thanks,
Louis
--
Dr Louis Rilling                        Kerlabs
Skype: louis.rilling                    Batiment Germanium
Phone: (+33|0) 6 80 89 08 23            80 avenue des Buttes de Coesmes
http://www.kerlabs.com/                 35700 Rennes
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.9 (GNU/Linux)
iEYEARECAAYFAk3k150ACgkQVKcRuvQ9Q1TiwQCeIhwpBR4ac6A7mpbE/TAA1ogP
MnkAoMm1TdieNdbaFdcO3kZWVD9XvqyI
=3Xhw
-----END PGP SIGNATURE-----