Re: LBA recorder disk speed tests
- Contemporary messages sorted: [ by date ] [ by thread ] [ by subject ] [ by author ] [ by messages with attachments ]
From: <Jamie.Stevens_at_email.protected>
Date: Tue, 25 Aug 2009 17:39:09 +1000
Hi Randall,
Even with a complete xfs format, it was taking only a few seconds for
our 2.5TB disks (although this wasn't your point).
cheers
Jamie
Randall Wayth wrote:
>Hi all,
>
>I think xfs is journaled, and I don't think you can disable the journal
>(at least from reading the manpage). It does give you the option of
>running the journal on a separate disk, which is quite nice. Perhaps it
>might be worth trying an unclean unmount with a relatively full xfs disk
>and see how long it takes to recover.
>
>The other nice thing about XFS is that you can delete lots of large
>files in a snap, so you don't actually need to "reformat" the disk to
>clean out old data. This might help with the scripty things Claire is
>talking about.
>
>-Randall.
>
>
>Cormac Reynolds wrote:
>>Nice work Jamie.
>>
>>Judging by these results it does look like xfs would probably be the way to go.
>>
>>Of course ext4 will be available pretty soon and should address the
>>performance problems in ext3, but for now xfs does seem the best
>>option.
>>
>>cheers,
>>Cormac.
>>
>>2009/8/25 Chris Phillips <Chris.Phillips_at_csiro.<!--nospam-->au>:
>>
>>>Yes but the problem with ext2 is that if the disk is uncleanly unmounted you
>>>have to wait ages to run fsync.
>>>
>>>
>>>xfs seems to be the speed of ext2 and not the problems. Reformatting changes
>>>are a trivial change to the scripts.
>>>
>>>
>>>Cheers
>>>Chris
>>>
>>>On 25/08/2009, at 4:12 PM, Claire Hotan wrote:
>>>
>>>
>>>>Hi all,
>>>>
>>>>Not necessarily relevant but Randall felt that xfs and raiding didn't get
>>>>on as well as ext2, so our internal feelings here were to simply turn off
>>>>the journalling option in the formatting script, thus returning things to
>>>>ext2, which means no significant changes but a probable potential speed up.
>>>>
>>>>I'd be interested to know which disk sets Jamie used, were they
>>>>"reliable"? The older disks seem more inclined to producing bigbuff errors
>>>>at high speed, though there's also been problems with Curtin 750s. In
>>>>general ATNF or CURT 500s that haven't had any history of disk failures seem
>>>>to be best able to cope with high speeds, but if we have some disk sets that
>>>>are marginal, then maybe switching to ext2 would fix those problems.
>>>>
>>>>Of course if people want to use xfs that's cool too, but from a scripts
>>>>perspective switching to ext2 would involve virtually no work, and we know
>>>>everything works just fine with ext2 because it's what we used until a year
>>>>ago.
>>>>
>>>>Just my two bob.
>>>>
>>>>cheers
>>>>
>>>>Claire
>>>>
>>>>
>>>>-----Original Message-----
>>>>From: owner-vlbiobs_at_atnf.<!--nospam-->csiro.au on behalf of Chris Phillips
>>>>Sent: Tue 8/25/2009 1:58 PM
>>>>To: Jamie Stevens
>>>>Cc: VLBI observers
>>>>Subject: Re: LBA recorder disk speed tests
>>>>
>>>>Thanks for doing this
>>>>
>>>>Of course these results show *any* of the filesystems can cope with
>>>>512 Mbps....
>>>>
>>>>Maybe we should try xfs for the next session. First we would need to
>>>>try some tests such as swapping disks after recording etc.
>>>>
>>>>Cheers
>>>>Chris
>>>>
>>>>On 25/08/2009, at 3:48 PM, Jamie Stevens wrote:
>>>>
>>>>
>Hi all,
>
>I've completed some diskspeed tests on cavsi1 using ext3, ext2 and xfs
>filesystems. I did all tests using filesizes that would have been
>generated by an experiment running at 512 Mbit/s; some with 10s files,
>others with 60s files.
>
>I've attached two postscript files to this email. Each shows the
>results
>of the 9 tests I did, and each file corresponds to tests run on one
>side
>of the Xraid (all tests were performed on both Xraid sides to ensure
>consistency).
>
>The results show that ext3 is quite a lot slower than ext2 or xfs,
>even
>when tuned for performance. xfs has the highest average speed, with
>ext2
>not far behind. It doesn't appear that changing the filesize has much
>effect on the average speed the disks can sustain.
>
>One of the most important things is the minimum speed. The bigbuffer
>will start to fill if the speed falls below the required recording
>rate,
>and this will lead to bigbuf skips if the disks don't catch up in
>time.
>ext3 filesystems have a very low minimum speed, as do ext2 filesystems
>with 10s files. But with 60s files, both ext2 and xfs do not drop much
>below 500 Mbit/s, and xfs can keep its minimum rate up even with 10s
>files.
>
>cheers
>Jamie
>
><
>lba_speedtests_xraid0
>.ps.gz><lba_speedtests_xraid1.ps.gz><Jamie_Stevens.vcf>
>
>>>>------------------------------------------------------------
>>>>Chris Phillips
>>>>CSIRO ATNF eVLBI project scientist
>>>>Office: (+61) (0)2 93724608 Mobile: (+61) (0)439487601
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>------------------------------------------------------------
>>>Chris Phillips
>>>CSIRO ATNF eVLBI project scientist
>>>Office: (+61) (0)2 93724608 Mobile: (+61) (0)439487601
>>>
>>>
>>>
>>>
>>>
>>
>>
>>
<div>
Received on 2009-08-25 17:39:46
Date: Tue, 25 Aug 2009 17:39:09 +1000
Hi Randall,
Even with a complete xfs format, it was taking only a few seconds for
our 2.5TB disks (although this wasn't your point).
cheers
Jamie
Randall Wayth wrote:
>Hi all,
>
>I think xfs is journaled, and I don't think you can disable the journal
>(at least from reading the manpage). It does give you the option of
>running the journal on a separate disk, which is quite nice. Perhaps it
>might be worth trying an unclean unmount with a relatively full xfs disk
>and see how long it takes to recover.
>
>The other nice thing about XFS is that you can delete lots of large
>files in a snap, so you don't actually need to "reformat" the disk to
>clean out old data. This might help with the scripty things Claire is
>talking about.
>
>-Randall.
>
>
>Cormac Reynolds wrote:
>>Nice work Jamie.
>>
>>Judging by these results it does look like xfs would probably be the way to go.
>>
>>Of course ext4 will be available pretty soon and should address the
>>performance problems in ext3, but for now xfs does seem the best
>>option.
>>
>>cheers,
>>Cormac.
>>
>>2009/8/25 Chris Phillips <Chris.Phillips_at_csiro.<!--nospam-->au>:
>>
>>>Yes but the problem with ext2 is that if the disk is uncleanly unmounted you
>>>have to wait ages to run fsync.
>>>
>>>
>>>xfs seems to be the speed of ext2 and not the problems. Reformatting changes
>>>are a trivial change to the scripts.
>>>
>>>
>>>Cheers
>>>Chris
>>>
>>>On 25/08/2009, at 4:12 PM, Claire Hotan wrote:
>>>
>>>
>>>>Hi all,
>>>>
>>>>Not necessarily relevant but Randall felt that xfs and raiding didn't get
>>>>on as well as ext2, so our internal feelings here were to simply turn off
>>>>the journalling option in the formatting script, thus returning things to
>>>>ext2, which means no significant changes but a probable potential speed up.
>>>>
>>>>I'd be interested to know which disk sets Jamie used, were they
>>>>"reliable"? The older disks seem more inclined to producing bigbuff errors
>>>>at high speed, though there's also been problems with Curtin 750s. In
>>>>general ATNF or CURT 500s that haven't had any history of disk failures seem
>>>>to be best able to cope with high speeds, but if we have some disk sets that
>>>>are marginal, then maybe switching to ext2 would fix those problems.
>>>>
>>>>Of course if people want to use xfs that's cool too, but from a scripts
>>>>perspective switching to ext2 would involve virtually no work, and we know
>>>>everything works just fine with ext2 because it's what we used until a year
>>>>ago.
>>>>
>>>>Just my two bob.
>>>>
>>>>cheers
>>>>
>>>>Claire
>>>>
>>>>
>>>>-----Original Message-----
>>>>From: owner-vlbiobs_at_atnf.<!--nospam-->csiro.au on behalf of Chris Phillips
>>>>Sent: Tue 8/25/2009 1:58 PM
>>>>To: Jamie Stevens
>>>>Cc: VLBI observers
>>>>Subject: Re: LBA recorder disk speed tests
>>>>
>>>>Thanks for doing this
>>>>
>>>>Of course these results show *any* of the filesystems can cope with
>>>>512 Mbps....
>>>>
>>>>Maybe we should try xfs for the next session. First we would need to
>>>>try some tests such as swapping disks after recording etc.
>>>>
>>>>Cheers
>>>>Chris
>>>>
>>>>On 25/08/2009, at 3:48 PM, Jamie Stevens wrote:
>>>>
>>>>
>Hi all,
>
>I've completed some diskspeed tests on cavsi1 using ext3, ext2 and xfs
>filesystems. I did all tests using filesizes that would have been
>generated by an experiment running at 512 Mbit/s; some with 10s files,
>others with 60s files.
>
>I've attached two postscript files to this email. Each shows the
>results
>of the 9 tests I did, and each file corresponds to tests run on one
>side
>of the Xraid (all tests were performed on both Xraid sides to ensure
>consistency).
>
>The results show that ext3 is quite a lot slower than ext2 or xfs,
>even
>when tuned for performance. xfs has the highest average speed, with
>ext2
>not far behind. It doesn't appear that changing the filesize has much
>effect on the average speed the disks can sustain.
>
>One of the most important things is the minimum speed. The bigbuffer
>will start to fill if the speed falls below the required recording
>rate,
>and this will lead to bigbuf skips if the disks don't catch up in
>time.
>ext3 filesystems have a very low minimum speed, as do ext2 filesystems
>with 10s files. But with 60s files, both ext2 and xfs do not drop much
>below 500 Mbit/s, and xfs can keep its minimum rate up even with 10s
>files.
>
>cheers
>Jamie
>
><
>lba_speedtests_xraid0
>.ps.gz><lba_speedtests_xraid1.ps.gz><Jamie_Stevens.vcf>
>
>>>>------------------------------------------------------------
>>>>Chris Phillips
>>>>CSIRO ATNF eVLBI project scientist
>>>>Office: (+61) (0)2 93724608 Mobile: (+61) (0)439487601
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>------------------------------------------------------------
>>>Chris Phillips
>>>CSIRO ATNF eVLBI project scientist
>>>Office: (+61) (0)2 93724608 Mobile: (+61) (0)439487601
>>>
>>>
>>>
>>>
>>>
>>
>>
>>
<div>
- text/x-vcard attachment: Jamie_Stevens.vcf