cancel
Showing results for 
Search instead for 
Did you mean: 

XEN and es/implementation

Former Member
0 Kudos

Note 941735 - SAP memory management for 64-bit Linux systems states, that using the std implementation can be used on Linux x86-64 and I heard rumours, that using that memory allocation method is 10 - 15 % faster than mmap.

I remember I read some note, that on a Xen PV instance "map" should be used, that note seems to be vanished (or that sentence)...

Is in that area anything recommended? Is memory disclaiming supported for mmap also?

Markus

Accepted Solutions (1)

Accepted Solutions (1)

nelis
Active Contributor
0 Kudos

[Note 962334|https://service.sap.com/sap/support/notes/962334] seems to be the central note for Xen but doesn't mention anything regarding es/implementation. Not sure if you've seen it ?

Nelis

markus_doehr2
Active Contributor
0 Kudos

Hi Nelis,

yes - I know... I just read "somewhere" that using map is faster than using std - I just can´t remember where... (blog? wiki? Hmm..)

Markus

nelis
Active Contributor
0 Kudos

Hi Marcus,

I did a quick search through my collection of notes on virtualization and [note 1122388|https://service.sap.com/sap/support/notes/1122388] on VMWare ESX suggests to use 'map' memory model for better memory throughput.

Regards,

Nelis

markus_doehr2
Active Contributor
0 Kudos

Thanx - that was the note where I read that!

I wonder if this is true for Xen also (which is a different hypervisor)...

Markus

hannes_kuehnemund
Active Contributor
0 Kudos

Hi Markus,

since SLES10 SP2 and RHEL5.2, both distributors deliver Xen patches that speed up memory accesses for SAP using es/implementation = std. Since then, the memory performance of Xen is almost at the level of native. Before these version, which means SP1/U1 and older, es/implementation = map was the way to go (with a cap of around 20%). The details on this are on the xen-devel mailing list. Just google for "xen-devel mprotect" and take the first hit.

Best Regards,

Hannes

markus_doehr2
Active Contributor
0 Kudos

Hi Hannes,

thanx very much! So memory disclaiming should now also work

Markus

Answers (0)