I'm sending this message to both dovecot and ceph-users ML so please don't mind if something seems too obvious for you.
Hi,
I have a question for both dovecot and ceph lists and below I'll explain what's going on.
Regarding dbox format (https://wiki2.dovecot.org/MailboxFormat/dbox), when using sdbox, a new file is stored for each email message. When using mdbox, multiple messages are appended to a single file until it reaches/passes the rotate limit.
I would like to understand better how the mdbox format impacts on IO performance. I think it's generally expected that fewer larger file translate to less IO and more troughput when compared to more small files, but how does dovecot handle that with mdbox? If dovecot does flush data to storage upon each and every new email is arrived and appended to the corresponding file, would that mean that it generate the same ammount of IO as it would do with one file per message? Also, if using mdbox many messages will be appended to a said file before a new file is created. That should mean that a file descriptor is kept open for sometime by dovecot process. Using cephfs as backend, how would this impact cluster performance regarding MDS caps and inodes cached when files from thousands of users are opened and appended all over?
I would like to understand this better.
Why? We are a small Business Email Hosting provider with bare metal, self hosted systems, using dovecot for servicing mailboxes and cephfs for email storage.
We are currently working on dovecot and storage redesign to be in production ASAP. The main objective is to serve more users with better performance, high availability and scalability.
- high availability and load balancing is extremely important to us *
On our current model, we're using mdbox format with dovecot, having dovecot's INDEXes stored in a replicated pool of SSDs, and messages stored in a replicated pool of HDDs (under a Cache Tier with a pool of SSDs). All using cephfs / filestore backend.
Currently there are 3 clusters running dovecot 2.2.34 and ceph Jewel (10.2.9-4).
- ~25K users from a few thousands of domains per cluster
- ~25TB of email data per cluster
- ~70GB of dovecot INDEX [meta]data per cluster
- ~100MB of cephfs metadata per cluster
Our goal is to build a single ceph cluster for storage that could expand in capacity, be highly available and perform well enough. I know, that's what everyone wants.
Cephfs is an important choise because:
- there can be multiple mountpoints, thus multiple dovecot instances on different hosts
- the same storage backend is used for all dovecot instances
- no need of sharding domains
- dovecot is easily load balanced (with director sticking users to the same dovecot backend)
On the upcoming upgrade we intent to:
- upgrade ceph to 12.X (Luminous)
- drop the SSD Cache Tier (because it's deprecated)
- use bluestore engine
I was said on freenode/#dovecot that there are many cases where SDBOX would perform better with NFS sharing. In case of cephfs, at first, I wouldn't think that would be true because more files == more generated IO, but thinking about what I said in the beginning regarding sdbox vs mdbox that could be wrong.
Any thoughts will be highlt appreciated.
Regards,
Webert Lima DevOps Engineer at MAV Tecnologia *Belo Horizonte - Brasil* *IRC NICK - WebertRLZ*
Hi,
some time back we had similar discussions when we, as an email provider, discussed to move away from traditional NAS/NFS storage to Ceph.
The problem with POSIX file systems and dovecot is that e.g. with mdbox only around ~20% of the IO operations are READ/WRITE, the rest are metadata IOs. You will not change this with using CephFS since it will basically behave the same way as e.g. NFS.
We decided to develop librmb to store emails as objects directly in RADOS instead of CephFS. The project is still under development, so you should not use it in production, but you can try it to run a POC.
For more information check out my slides from Ceph Day London 2018: https://dalgaaf.github.io/cephday-london2018-emailstorage/#/cover-page
The project can be found on github: https://github.com/ceph-dovecot/
-Danny
Am 16.05.2018 um 20:37 schrieb Webert de Souza Lima:
I'm sending this message to both dovecot and ceph-users ML so please don't mind if something seems too obvious for you.
Hi,
I have a question for both dovecot and ceph lists and below I'll explain what's going on.
Regarding dbox format (https://wiki2.dovecot.org/MailboxFormat/dbox), when using sdbox, a new file is stored for each email message. When using mdbox, multiple messages are appended to a single file until it reaches/passes the rotate limit.
I would like to understand better how the mdbox format impacts on IO performance. I think it's generally expected that fewer larger file translate to less IO and more troughput when compared to more small files, but how does dovecot handle that with mdbox? If dovecot does flush data to storage upon each and every new email is arrived and appended to the corresponding file, would that mean that it generate the same ammount of IO as it would do with one file per message? Also, if using mdbox many messages will be appended to a said file before a new file is created. That should mean that a file descriptor is kept open for sometime by dovecot process. Using cephfs as backend, how would this impact cluster performance regarding MDS caps and inodes cached when files from thousands of users are opened and appended all over?
I would like to understand this better.
Why? We are a small Business Email Hosting provider with bare metal, self hosted systems, using dovecot for servicing mailboxes and cephfs for email storage.
We are currently working on dovecot and storage redesign to be in production ASAP. The main objective is to serve more users with better performance, high availability and scalability.
- high availability and load balancing is extremely important to us *
On our current model, we're using mdbox format with dovecot, having dovecot's INDEXes stored in a replicated pool of SSDs, and messages stored in a replicated pool of HDDs (under a Cache Tier with a pool of SSDs). All using cephfs / filestore backend.
Currently there are 3 clusters running dovecot 2.2.34 and ceph Jewel (10.2.9-4).
- ~25K users from a few thousands of domains per cluster
- ~25TB of email data per cluster
- ~70GB of dovecot INDEX [meta]data per cluster
- ~100MB of cephfs metadata per cluster
Our goal is to build a single ceph cluster for storage that could expand in capacity, be highly available and perform well enough. I know, that's what everyone wants.
Cephfs is an important choise because:
- there can be multiple mountpoints, thus multiple dovecot instances on different hosts
- the same storage backend is used for all dovecot instances
- no need of sharding domains
- dovecot is easily load balanced (with director sticking users to the same dovecot backend)
On the upcoming upgrade we intent to:
- upgrade ceph to 12.X (Luminous)
- drop the SSD Cache Tier (because it's deprecated)
- use bluestore engine
I was said on freenode/#dovecot that there are many cases where SDBOX would perform better with NFS sharing. In case of cephfs, at first, I wouldn't think that would be true because more files == more generated IO, but thinking about what I said in the beginning regarding sdbox vs mdbox that could be wrong.
Any thoughts will be highlt appreciated.
Regards,
Webert Lima DevOps Engineer at MAV Tecnologia *Belo Horizonte - Brasil* *IRC NICK - WebertRLZ*
ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
Hello Danny,
I actually saw that thread and I was very excited about it. I thank you all for that idea and all the effort being put in it. I haven't yet tried to play around with your plugin but I intend to, and to contribute back. I think when it's ready for production it will be unbeatable.
I have watched your talk at Cephalocon (on YouTube). I'll see your slides, maybe they'll give me more insights on our infrastructure architecture.
As you can see our business is still taking baby steps compared to Deutsche Telekom's but we face infrastructure challenges everyday since ever. As for now, I think we could still fit with cephfs, but we definitely need some improvement.
Regards,
Webert Lima DevOps Engineer at MAV Tecnologia *Belo Horizonte - Brasil* *IRC NICK - WebertRLZ*
On Wed, May 16, 2018 at 4:42 PM Danny Al-Gaaf <danny.al-gaaf@bisect.de> wrote:
Hi,
some time back we had similar discussions when we, as an email provider, discussed to move away from traditional NAS/NFS storage to Ceph.
The problem with POSIX file systems and dovecot is that e.g. with mdbox only around ~20% of the IO operations are READ/WRITE, the rest are metadata IOs. You will not change this with using CephFS since it will basically behave the same way as e.g. NFS.
We decided to develop librmb to store emails as objects directly in RADOS instead of CephFS. The project is still under development, so you should not use it in production, but you can try it to run a POC.
For more information check out my slides from Ceph Day London 2018: https://dalgaaf.github.io/cephday-london2018-emailstorage/#/cover-page
The project can be found on github: https://github.com/ceph-dovecot/
-Danny
Am 16.05.2018 um 20:37 schrieb Webert de Souza Lima:
I'm sending this message to both dovecot and ceph-users ML so please don't mind if something seems too obvious for you.
Hi,
I have a question for both dovecot and ceph lists and below I'll explain what's going on.
Regarding dbox format (https://wiki2.dovecot.org/MailboxFormat/dbox), when using sdbox, a new file is stored for each email message. When using mdbox, multiple messages are appended to a single file until it reaches/passes the rotate limit.
I would like to understand better how the mdbox format impacts on IO performance. I think it's generally expected that fewer larger file translate to less IO and more troughput when compared to more small files, but how does dovecot handle that with mdbox? If dovecot does flush data to storage upon each and every new email is arrived and appended to the corresponding file, would that mean that it generate the same ammount of IO as it would do with one file per message? Also, if using mdbox many messages will be appended to a said file before a new file is created. That should mean that a file descriptor is kept open for sometime by dovecot process. Using cephfs as backend, how would this impact cluster performance regarding MDS caps and inodes cached when files from thousands of users are opened and appended all over?
I would like to understand this better.
Why? We are a small Business Email Hosting provider with bare metal, self hosted systems, using dovecot for servicing mailboxes and cephfs for email storage.
We are currently working on dovecot and storage redesign to be in production ASAP. The main objective is to serve more users with better performance, high availability and scalability.
- high availability and load balancing is extremely important to us *
On our current model, we're using mdbox format with dovecot, having dovecot's INDEXes stored in a replicated pool of SSDs, and messages stored in a replicated pool of HDDs (under a Cache Tier with a pool of SSDs). All using cephfs / filestore backend.
Currently there are 3 clusters running dovecot 2.2.34 and ceph Jewel (10.2.9-4).
- ~25K users from a few thousands of domains per cluster
- ~25TB of email data per cluster
- ~70GB of dovecot INDEX [meta]data per cluster
- ~100MB of cephfs metadata per cluster
Our goal is to build a single ceph cluster for storage that could expand in capacity, be highly available and perform well enough. I know, that's what everyone wants.
Cephfs is an important choise because:
- there can be multiple mountpoints, thus multiple dovecot instances on different hosts
- the same storage backend is used for all dovecot instances
- no need of sharding domains
- dovecot is easily load balanced (with director sticking users to the same dovecot backend)
On the upcoming upgrade we intent to:
- upgrade ceph to 12.X (Luminous)
- drop the SSD Cache Tier (because it's deprecated)
- use bluestore engine
I was said on freenode/#dovecot that there are many cases where SDBOX would perform better with NFS sharing. In case of cephfs, at first, I wouldn't think that would be true because more files == more generated IO, but thinking about what I said in the beginning regarding sdbox vs mdbox that could be wrong.
Any thoughts will be highlt appreciated.
Regards,
Webert Lima DevOps Engineer at MAV Tecnologia *Belo Horizonte - Brasil* *IRC NICK - WebertRLZ*
ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
participants (2)
-
Danny Al-Gaaf
-
Webert de Souza Lima