Proper configuration of the SSDs in a storage brick

Proper configuration of the SSDs in a storage brick

[Stephen Perkins]

Hi all,

In looking at the design of a storage brick (just OSDs), I have found a dual
power hardware solution that allows for 10 hot-swap drives and has a
motherboard with 2 SATA III 6G ports (for the SSDs) and 8 SATA II 3G (for
physical drives).  No RAID card. This seems a good match to me given my
needs.  This system also supports 10G Ethernet via an add in card, so please
assume that for the questions.  I'm also assuming 2TB or 3TB drives for the
8 hot swap.  My workload is throughput intensive (writes mainly) and not IOP
heavy.

I have 2 questions and would love to hear from the group.

Question 1: What is the most appropriate configuration for the journal SSDs?

I'm not entirely sure what happens when you lose a journal drive.  If the
whole brick goes offline (i.e. all OSDs stop communicating with ceph), does
it make since to configure the SSDs into RAID1?

Alternatively, it seems that there is a performance benefit to having 2
independent SSDs since you get potentially twice the journal rate.  If a
journal drive goes offline. do you only have to recover half the brick?

If having 2 drives does not provide a performance benefit, it there a
benefit other than RAID 1 for redundancy?


Question 2:  How to handle the OS?

I need to install an OS on each brick?   I'm guessing the SSDs are the
device of choice. Not being entirely familiar with the journal drives:

Should I create a separate drive partition for the OS?  

Or. can the journals write to the same partition as the OS?  

Should I dedicate one drive to the OS and one drive to the journal?

RAID1 or independent?

Use a mechanical drive?

Alternately. the 10G NIC cards support remote iSCSI boot.  This allows both
SSDs to be dedicated to journaling. Seems like more complexity. 

I would appreciate hearing the thoughts of the group.

Best regards,

- Steve

Re: Proper configuration of the SSDs in a storage brick

[Wido den Hollander]

On 10/25/2012 03:30 PM, Stephen Perkins wrote:
> Hi all,
>
> In looking at the design of a storage brick (just OSDs), I have found a dual
> power hardware solution that allows for 10 hot-swap drives and has a
> motherboard with 2 SATA III 6G ports (for the SSDs) and 8 SATA II 3G (for
> physical drives).  No RAID card. This seems a good match to me given my
> needs.  This system also supports 10G Ethernet via an add in card, so please
> assume that for the questions.  I'm also assuming 2TB or 3TB drives for the
> 8 hot swap.  My workload is throughput intensive (writes mainly) and not IOP
> heavy.
>
> I have 2 questions and would love to hear from the group.
>
> Question 1: What is the most appropriate configuration for the journal SSDs?
>
> I'm not entirely sure what happens when you lose a journal drive.  If the
> whole brick goes offline (i.e. all OSDs stop communicating with ceph), does
> it make since to configure the SSDs into RAID1?
>

When you loose the journal these OSDs will commit suicide and in this 
case you'd loose 8 OSDs.

Placing two SSDs in RAID-1 seems like overkill to me. I've been using 
hundreds of Intel SSDs over the past 3 years and I've never see one (not 
one!) die.

A SSD will die at some point due to extensive writes, but in RAID-1 they 
would burn through those writes in a identical matter.

> Alternatively, it seems that there is a performance benefit to having 2
> independent SSDs since you get potentially twice the journal rate.  If a
> journal drive goes offline. do you only have to recover half the brick?
>

If you place 4 OSDs on 1 SSD and the other 4 on the second SSD you'd 
indeed only loose 4 OSDs.

> If having 2 drives does not provide a performance benefit, it there a
> benefit other than RAID 1 for redundancy?
>

Something like RAID-1 would not, RAID-0 might do it. But I would split 
the OSDs up over 2 SSDs.

>
> Question 2:  How to handle the OS?
>
> I need to install an OS on each brick?   I'm guessing the SSDs are the
> device of choice. Not being entirely familiar with the journal drives:
>
> Should I create a separate drive partition for the OS?
>
> Or. can the journals write to the same partition as the OS?
>
> Should I dedicate one drive to the OS and one drive to the journal?
>

I'd suggest using Intel SSDs and shrinking them in size using HPA, Host 
Protected Area.

With that you can shrinkg a 180GB SSD to for example 60GB. By doing so 
the SSD can perform better wear-leveling and it would maintain optimal 
performance over time, it also extends the lifetime of the SSD. It has 
more "spare cells".

Under Linux you can change this with "hdparm" and the -N option.

Using a separate partition for the journal and OS would be preferred. 
Make sure to align the partition with the erase size of the SSD, 
otherwise you could run into write amplification of the SSD.

You would end up with:
* OS partition
* Swap?
* Journal #1
* Journal #2

Depends on what you are going to use.

Wido

> RAID1 or independent?
>
> Use a mechanical drive?
>
> Alternately. the 10G NIC cards support remote iSCSI boot.  This allows both
> SSDs to be dedicated to journaling. Seems like more complexity.
>
> I would appreciate hearing the thoughts of the group.
>
> Best regards,
>
> - Steve
>
>
>
> --
> To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>

RE: Proper configuration of the SSDs in a storage brick

[Stephen Perkins]

Most excellent!  Many thanks for the clarification.  Questions:

>  Something like RAID-1 would not, RAID-0 might do it. But I would split
the OSDs up over 2 SSDs.

I could take a 256G SSD and then use 50% which gives me 128G:
	16G for OS / SWAP (Assume 24GB RAM -> 2G per OSD plus 8G for
OS/Swap)
	8 * 15G journal

Q1:
	 Is a 15G journal large enough?

Q2: 
	Given an approximate max theoretical of 500-600 MB/s sustained
throughput of SSD
	 (I am throughput intensive) and 10G Ethernet... do I need 2 SSDs
for performance or
	 will one do?

(Given a theoretical mechanical drive throughput is (100->125 MB/s * 8) > a
single SSD).

-Steve


-----Original Message-----
From: Wido den Hollander [mailto:wido@widodh.nl] 
Sent: Friday, October 26, 2012 8:56 AM
To: Stephen Perkins
Cc: ceph-devel@vger.kernel.org
Subject: Re: Proper configuration of the SSDs in a storage brick

On 10/25/2012 03:30 PM, Stephen Perkins wrote:
> Hi all,
>
> In looking at the design of a storage brick (just OSDs), I have found 
> a dual power hardware solution that allows for 10 hot-swap drives and 
> has a motherboard with 2 SATA III 6G ports (for the SSDs) and 8 SATA 
> II 3G (for physical drives).  No RAID card. This seems a good match to 
> me given my needs.  This system also supports 10G Ethernet via an add 
> in card, so please assume that for the questions.  I'm also assuming 
> 2TB or 3TB drives for the
> 8 hot swap.  My workload is throughput intensive (writes mainly) and 
> not IOP heavy.
>
> I have 2 questions and would love to hear from the group.
>
> Question 1: What is the most appropriate configuration for the journal
SSDs?
>
> I'm not entirely sure what happens when you lose a journal drive.  If 
> the whole brick goes offline (i.e. all OSDs stop communicating with 
> ceph), does it make since to configure the SSDs into RAID1?
>

When you loose the journal these OSDs will commit suicide and in this case
you'd loose 8 OSDs.

Placing two SSDs in RAID-1 seems like overkill to me. I've been using
hundreds of Intel SSDs over the past 3 years and I've never see one (not
one!) die.

A SSD will die at some point due to extensive writes, but in RAID-1 they
would burn through those writes in a identical matter.

> Alternatively, it seems that there is a performance benefit to having 
> 2 independent SSDs since you get potentially twice the journal rate.  
> If a journal drive goes offline. do you only have to recover half the
brick?
>

If you place 4 OSDs on 1 SSD and the other 4 on the second SSD you'd indeed
only loose 4 OSDs.

> If having 2 drives does not provide a performance benefit, it there a 
> benefit other than RAID 1 for redundancy?
>

Something like RAID-1 would not, RAID-0 might do it. But I would split the
OSDs up over 2 SSDs.

>
> Question 2:  How to handle the OS?
>
> I need to install an OS on each brick?   I'm guessing the SSDs are the
> device of choice. Not being entirely familiar with the journal drives:
>
> Should I create a separate drive partition for the OS?
>
> Or. can the journals write to the same partition as the OS?
>
> Should I dedicate one drive to the OS and one drive to the journal?
>

I'd suggest using Intel SSDs and shrinking them in size using HPA, Host
Protected Area.

With that you can shrinkg a 180GB SSD to for example 60GB. By doing so the
SSD can perform better wear-leveling and it would maintain optimal
performance over time, it also extends the lifetime of the SSD. It has more
"spare cells".

Under Linux you can change this with "hdparm" and the -N option.

Using a separate partition for the journal and OS would be preferred. 
Make sure to align the partition with the erase size of the SSD, otherwise
you could run into write amplification of the SSD.

You would end up with:
* OS partition
* Swap?
* Journal #1
* Journal #2

Depends on what you are going to use.

Wido

> RAID1 or independent?
>
> Use a mechanical drive?
>
> Alternately. the 10G NIC cards support remote iSCSI boot.  This allows 
> both SSDs to be dedicated to journaling. Seems like more complexity.
>
> I would appreciate hearing the thoughts of the group.
>
> Best regards,
>
> - Steve
>
>
>
> --
> To unsubscribe from this list: send the line "unsubscribe ceph-devel" 
> in the body of a message to majordomo@vger.kernel.org More majordomo 
> info at  http://vger.kernel.org/majordomo-info.html
>

Re: Proper configuration of the SSDs in a storage brick

[Sage Weil]

On Fri, 26 Oct 2012, Wido den Hollander wrote:
> On 10/25/2012 03:30 PM, Stephen Perkins wrote:
> > Hi all,
> > 
> > In looking at the design of a storage brick (just OSDs), I have found a dual
> > power hardware solution that allows for 10 hot-swap drives and has a
> > motherboard with 2 SATA III 6G ports (for the SSDs) and 8 SATA II 3G (for
> > physical drives).  No RAID card. This seems a good match to me given my
> > needs.  This system also supports 10G Ethernet via an add in card, so please
> > assume that for the questions.  I'm also assuming 2TB or 3TB drives for the
> > 8 hot swap.  My workload is throughput intensive (writes mainly) and not IOP
> > heavy.
> > 
> > I have 2 questions and would love to hear from the group.
> > 
> > Question 1: What is the most appropriate configuration for the journal SSDs?
> > 
> > I'm not entirely sure what happens when you lose a journal drive.  If the
> > whole brick goes offline (i.e. all OSDs stop communicating with ceph), does
> > it make since to configure the SSDs into RAID1?
> > 
> 
> When you loose the journal these OSDs will commit suicide and in this case
> you'd loose 8 OSDs.

One small correction here: it depends.

If you use ext4 or XFS, then yes: losing the journal means the data disk 
is lost too.  

However, if you use btrfs, the data disk has consistent point-in-time 
checkpoints it can roll back to.  It is not useful from the perspective of 
a specific IO request (i.e., if client X wrote to 3 replicas, then all 3 
replicas lost their journals, the write may have been lost, along with the 
other 0.01% of writes that happened in the last several seconds).

On the other hand, for a single osd that loses the journal, you can 
initialize a new one and it can rejoin the cluster and will have 99.99% of 
the data in place, making reintegration/recovery quick and cheap

sage

Re: Proper configuration of the SSDs in a storage brick

[Gregory Farnum]

On Fri, Oct 26, 2012 at 7:17 AM, Stephen Perkins <perkins@netmass.com> wrote:
> Most excellent!  Many thanks for the clarification.  Questions:
>
>>  Something like RAID-1 would not, RAID-0 might do it. But I would split
> the OSDs up over 2 SSDs.
>
> I could take a 256G SSD and then use 50% which gives me 128G:
>         16G for OS / SWAP (Assume 24GB RAM -> 2G per OSD plus 8G for
> OS/Swap)
>         8 * 15G journal
>
> Q1:
>          Is a 15G journal large enough?

Our rule of thumb is that your journal should be able to absorb all
writes coming into the OSD for a period of 10-20 seconds. Given 10GbE
and 8 OSDs, you're looking at ~125MB/s per OSD, so a 15GB journal
should be good.


> Q2:
>         Given an approximate max theoretical of 500-600 MB/s sustained
> throughput of SSD
>          (I am throughput intensive) and 10G Ethernet... do I need 2 SSDs
> for performance or
>          will one do?
>
> (Given a theoretical mechanical drive throughput is (100->125 MB/s * 8) > a
> single SSD).

Sounds like you need 2 SSDs, then!
-Greg

>
> -Steve
>
>
> -----Original Message-----
> From: Wido den Hollander [mailto:wido@widodh.nl]
> Sent: Friday, October 26, 2012 8:56 AM
> To: Stephen Perkins
> Cc: ceph-devel@vger.kernel.org
> Subject: Re: Proper configuration of the SSDs in a storage brick
>
> On 10/25/2012 03:30 PM, Stephen Perkins wrote:
>> Hi all,
>>
>> In looking at the design of a storage brick (just OSDs), I have found
>> a dual power hardware solution that allows for 10 hot-swap drives and
>> has a motherboard with 2 SATA III 6G ports (for the SSDs) and 8 SATA
>> II 3G (for physical drives).  No RAID card. This seems a good match to
>> me given my needs.  This system also supports 10G Ethernet via an add
>> in card, so please assume that for the questions.  I'm also assuming
>> 2TB or 3TB drives for the
>> 8 hot swap.  My workload is throughput intensive (writes mainly) and
>> not IOP heavy.
>>
>> I have 2 questions and would love to hear from the group.
>>
>> Question 1: What is the most appropriate configuration for the journal
> SSDs?
>>
>> I'm not entirely sure what happens when you lose a journal drive.  If
>> the whole brick goes offline (i.e. all OSDs stop communicating with
>> ceph), does it make since to configure the SSDs into RAID1?
>>
>
> When you loose the journal these OSDs will commit suicide and in this case
> you'd loose 8 OSDs.
>
> Placing two SSDs in RAID-1 seems like overkill to me. I've been using
> hundreds of Intel SSDs over the past 3 years and I've never see one (not
> one!) die.
>
> A SSD will die at some point due to extensive writes, but in RAID-1 they
> would burn through those writes in a identical matter.
>
>> Alternatively, it seems that there is a performance benefit to having
>> 2 independent SSDs since you get potentially twice the journal rate.
>> If a journal drive goes offline. do you only have to recover half the
> brick?
>>
>
> If you place 4 OSDs on 1 SSD and the other 4 on the second SSD you'd indeed
> only loose 4 OSDs.
>
>> If having 2 drives does not provide a performance benefit, it there a
>> benefit other than RAID 1 for redundancy?
>>
>
> Something like RAID-1 would not, RAID-0 might do it. But I would split the
> OSDs up over 2 SSDs.
>
>>
>> Question 2:  How to handle the OS?
>>
>> I need to install an OS on each brick?   I'm guessing the SSDs are the
>> device of choice. Not being entirely familiar with the journal drives:
>>
>> Should I create a separate drive partition for the OS?
>>
>> Or. can the journals write to the same partition as the OS?
>>
>> Should I dedicate one drive to the OS and one drive to the journal?
>>
>
> I'd suggest using Intel SSDs and shrinking them in size using HPA, Host
> Protected Area.
>
> With that you can shrinkg a 180GB SSD to for example 60GB. By doing so the
> SSD can perform better wear-leveling and it would maintain optimal
> performance over time, it also extends the lifetime of the SSD. It has more
> "spare cells".
>
> Under Linux you can change this with "hdparm" and the -N option.
>
> Using a separate partition for the journal and OS would be preferred.
> Make sure to align the partition with the erase size of the SSD, otherwise
> you could run into write amplification of the SSD.
>
> You would end up with:
> * OS partition
> * Swap?
> * Journal #1
> * Journal #2
>
> Depends on what you are going to use.
>
> Wido
>
>> RAID1 or independent?
>>
>> Use a mechanical drive?
>>
>> Alternately. the 10G NIC cards support remote iSCSI boot.  This allows
>> both SSDs to be dedicated to journaling. Seems like more complexity.
>>
>> I would appreciate hearing the thoughts of the group.
>>
>> Best regards,
>>
>> - Steve
>>
>>
>>
>> --
>> To unsubscribe from this list: send the line "unsubscribe ceph-devel"
>> in the body of a message to majordomo@vger.kernel.org More majordomo
>> info at  http://vger.kernel.org/majordomo-info.html
>>
>
>
> --
> To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html