On 12/11/2013 12:59 AM, Duan, Jiangang wrote:
Thanks. I in general think to find one implementation suitable for all usage models
(small vs. big, cold vs. hot) is very difficult. So I like the idea of "a backend
that lets you plug in a next-gen backend beneath it - " -
K/V may be a good way to handle many small objects than XFS - however I am not
sure if levelDB is the right choice (consider it is good for write than read)
and also not sure K/V this will benefit RBD workload or not (consider all 4MB
object size).
Will think more about this and talk with you again.
I have been very interested in this topic recently and have been doing
some benchmarking with basho's leveldb, hyperdex, and stock leveldb
implementations. Each has certain advantages (usually related to
whatever they advertise, ie crc32 for basho, etc), but all seem to have
poor sync read/write performance, both sequential and random. I want to
take a look at rocksdb to see how it compares as well.
This whole area is very interesting as there are obvious tradeoffs
regarding how we do things now vs what we potentially could do down the
road. Being able to eliminate POSIX entirely behind the scenes would
obviously be nice for a lot of reasons.
Mark
-jiangang
-----Original Message-----
From: Sage Weil [mailto:s...@inktank.com]
Sent: Wednesday, December 11, 2013 2:09 PM
To: Duan, Jiangang
Cc: ceph-devel@vger.kernel.org
Subject: Re: question for the new ceph-osd key/value backend
Hi Jiangang,
On Wed, 11 Dec 2013, Duan, Jiangang wrote:
Sage,
I have some questions regarding to the key/value backend work.
What is the motivation to work on this? (or what is the problem we
want to solve?)
1) to use the new interface thus we can bypass all the OS layer thus get a
short latency?
That is one part. The current strategy of layering on top of a file system and
using a write-ahead journal makes sense given the existing linux fs building
blocks, but is far from an optimal solution for many workloads. A k/v
interface based on something leveldb probably performs much better for many
small-object use-cases. Also, a k/v backend can take advatange of emerging
non-block storage interfaces like NVMKV, Kinetic, new libraries like rocksdb,
etc.
2) or to leverage some new primitive e.g. the atomic write thus to simplify the
code writing?
That too. Basically, we are currently doing a lot of work to get what we need
out of posix, and are paying the price.
There are several different possibilities to use future NVM technology
- NVM.FILE, NVM.BLOCK, PM.XXX
http://snia.org/sites/default/files/NVMProgrammingModel_v1r10DRAFT.pdf
Even for openNVM thing - there are other usage model than k/v.
Do you have any typical usage model for this?
I wasn't familiar with these; thanks for the reference! Of these, NVM.FILE
seems the most interesting (it maps most closely to an object).
I am predisposed to skepticism when it comes to these sorts of standards/API
docs that precede an actual implementation, but it is encourgaging to see some
effort here towards a common interface.
In the end, we want to support generic Ceph workloads. These range from rbd
block and file type workloads (objects are stripes of files, with random bytes
rewritten) to omap type workloads (like rgw bucket indices that are purely
key/value).
I think the first wins would be:
1- a backend that more efficiently handles rgw bucket index workloads
2- a backend that is more efficient for rgw in general (i.e., immutable
objects)
3- a backend that can handle more general purpose workloads (like rbd and
cephfs)
and separately,
4- a backend that lets you plug in a next-gen backend beneath it, like NVMKV
and speedy flash.
sage
-jiangang
===================
From: Sage Weil <sage <at> inktank.com>
Subject: new ceph-osd key/value backend
Newsgroups: gmane.comp.file-systems.ceph.devel
Date: 2013-11-09 10:09:52 GMT (4 weeks, 3 days, 16 hours and 39
minutes ago) I've written up a blueprint with a rough sketch of how to
take advantage of alternative storage interfaces. I am very happy to
see that several f them have emerged over the past year or two:
- fusionio's KVMKV is a key/value interface for their flash products
- seagate's kinetic is a key/value interface for their new
ethernet-based drive
Also, leveldb is pretty great for many workloads when run on a
tranditional disk/fs.
The good news is a lot of the existing work that went into support
omap looks to be reusable here. Some new functionality and
refactoring is needed, though, particularly when it comes to storing
object data (the file-like bag of bytes portion) as key/value pairs.
The blueprint is here:
http://wiki.ceph.com/01Planning/02Blueprints/Firefly/osd%3A_new_key%2F
%2Fvalue_backend
N?????r??y??????X???v???)?{.n?????z?]z????ay?????j ??f???h??????w???
???j:+v???w???????? ????zZ+???????j"????i
N�����r��y���b�X��ǧv�^�){.n�+���z�]z�{ay�ʇڙ�,j��f���h���z��w���
���j:+v���w�j�m��������zZ+��ݢj"��!tml=
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html