Thx!
BTW,
We have thought about if it is a waste to put only 1 OSD on a whole NVMe
SSD, especially a high-end one, such as Intel DC P3700.
And we read some spdk codes, noticed that it is band with NVMe namespace.
Theoretically, a NVMe SSD can be multi-namespaces by partition LBAs.
Each namespace has its own Submission/Completion Queues, so it would work well.
But now, took P3700 as an example, it only supports NVMe v1.0, and the Namespace
Management Cmd Sets is not supported yet. It would be easier to create namespaces someday.
And "os/bluestore/NVMEDevices.cc" identifies 1 namespace per SSD too.
Hopes to see multi-namespaces support to make full use of SSDs.
--
Best Regards!
On 4/22/16, 20:02, "Haomai Wang" <haomai@xxxxxxxx> wrote:
>On Fri, Apr 22, 2016 at 5:52 AM, Li Tianxiang <lanceflee@xxxxxxx> wrote:
>>
>> Sorry,
>> I mean that we have two NVMe SSD each node,one for a OSD,. But only 1 OSD is up.
>>
>> It seams that the hugepages are all taken by 1 spdk instance, nothing left for another.
>
>Yep, this is probrem.....
>
>we have two options now, one is use dpdk multi process to let two
>ceph-osd aware, it need some codes. another is waiting for "fedora osd
>process" feature which is more suitable......
>
>>
>>
>>
>>
>> --
>> Best Regards!
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>> On 4/22/16, 17:28, "李 天祥" <ceph-devel-owner@xxxxxxxxxxxxxxx on behalf of lanceflee@xxxxxxx> wrote:
>>
>>>1. Fisrt of all , $ yum remove dpdk-devel dpdk -y
>>>2.get dpdk-2.2.0 source code , and compile it as shared library, As follows:
>>> a. vim config/common_linuxapp, modify "CONFIG_RTE_BUILD_SHARED_LIB=y"
>>> b.make install T=x86_64-native-linuxapp-gcc DESTDIR=/usr
>>>3.build ceph ./configure --with-spdk
>>
��.n��������+%������w��{.n����z��u���ܨ}���Ơz�j:+v�����w����ޙ��&�)ߡ�a����z�ޗ���ݢj��w�f
