[pve-devel] back from holiday, I'll more time to work on proxmox

[Alexandre DERUMIER]

>>Also client side, a new qemu optimisation has been done, reducing memory copy between qemu and librbd, so reduce latency.
>>
>>https://lists.gnu.org/archive/html/qemu-devel/2017-02/msg04545.html
>>
>>this patch is not yet upstream (I think it has missed the qemu 2.9 windows), but it's in good state.
>>I'll try to test it.

speaking too fast, the patch is already include in qemu 2.9

http://git.qemu-project.org/?p=qemu.git;a=commit;h=1d393bdeae22fde2cb83c1ea719675747c85c40e

It should give a nice boost for small iops with librbd.




----- Mail original -----
De: "aderumier" <aderumier at odiso.com>
À: "Fabian Grünbichler" <f.gruenbichler at proxmox.com>
Cc: "pve-devel" <pve-devel at pve.proxmox.com>, "dietmar" <dietmar at proxmox.com>
Envoyé: Mardi 18 Avril 2017 06:39:13
Objet: Re: [pve-devel] back from holiday, I'll more time to work on proxmox

>>do you have any hard figures regarding the jemalloc benefits? 

We have done benchmark with ceph guys last year, 

it's really reduce latency for low iodepth, and with a lot of clients I was jumping from 450k iops to 650k iops. 

(It's use less cpu). 

It's not enable by default, because it's use a little bit more memory by osd (I think around 500mb), and ceph guys didn't want impact customers 
with big box (like 48 osd). 


The patch is pretty simple, I'll test it again with liminous. 


Also client side, a new qemu optimisation has been done, reducing memory copy between qemu and librbd, so reduce latency. 

https://lists.gnu.org/archive/html/qemu-devel/2017-02/msg04545.html 

this patch is not yet upstream (I think it has missed the qemu 2.9 windows), but it's in good state. 
I'll try to test it. 




----- Mail original ----- 
De: "Fabian Grünbichler" <f.gruenbichler at proxmox.com> 
À: "Alexandre Derumier" <aderumier at odiso.com>, "pve-devel" <pve-devel at pve.proxmox.com>, "dietmar" <dietmar at proxmox.com> 
Envoyé: Lundi 17 Avril 2017 16:18:56 
Objet: Re: [pve-devel] back from holiday, I'll more time to work on proxmox 

> Alexandre DERUMIER <aderumier at odiso.com> hat am 16. April 2017 um 20:55 geschrieben: 
> 
> 
> >>Great. But keep in mind that Debian stretch release is soon, so it is too late 
> >>for big new features in 5.0. But we are a rolling release, so we can integrate 
> >>those 
> >>new feature when they become ready/stable. For 5.0, we are mostly in bug fixing 
> >>mode. New features which does not add much code are still accepted. 
> 
> Yes, no problem ! (I had hope to have more time before, but I was very very busy last months). 
> 
> 
> 
> >>We also have the a new ceph repository which requires more testing ... 
> 
> do you build ceph package yourself ? If yes, it could be great to build with jemalloc. 
> I can provide patch if needed. 

yes, see [1] - 12.x only (for now). the current patch set / delta is not that big yet. if you have specific tweaks you'd like to see, please file enhancement requests, propose them on pve-devel or just send patches (although I guess that most of them would either be cherry-picks or build options currently, where sending patches adding new patches does not make that much sense IMHO ;)). 

our main goal with this is to have better control over updates - both in the sense of being able to test our integration better before passing on an upstream release, and in the sense of being able to quickly fix issues without having to wait for them to be included in an upstream release. 

do you have any hard figures regarding the jemalloc benefits? 

1: https://git.proxmox.com/?p=ceph.git;a=summary