List:General Discussion« Previous MessageNext Message »
From:Chris Knipe Date:July 26 2013 10:58pm
Subject:Re: hypothetical question about data storage
View as plain text  
Well that information I can provide....

As mentioned, we use an md5 (hex) checksum to track the files.  In
terms of the tables, I would definately consider the md5 checksum as a
PK (char(32) due to the hex nature), a blob for the data, and then
there will also be a datetime column to indicate when last the file
was accessed.  We already use mySQL with 4096 innodb tables in a
single database to track the timestamp when the file was last accessed
(noatime on the file system) - and it's working remarkably well.  But
I do understand that adding the blob will most certainly change things
due to the fact that much more data needs to be moved arround (memory,
disk and networking layers).

Currently the server (32GB ram, dedicated 2 x quad core xeon) is
pretty much idling in terms of load, doing approximately 100
transactions per second with less than 50 out of a max of 500
connections configured (binlogs and all). Note however that this is
just inserting new records, and updating the last accessed timestamp
on the records.  Sometimes, there are large delete transactions
running as well to remove expired files (start transaction; delete
from... ; commit)

We store the files in 4096 unique directories, the files are
structured very simply as follows:

<1st char>/<1st + 2nd char>/<1st + 2nd + 3rd char>/<filename being
hex md5 checksum>

Currently, there is about 3TB worth of data, this -will- grow easily
to 10 times in size over time (200GB to 300GB per day).  Per 1TB of
data, we are basically looking at 1m files (or records) split on
average at 243 files per directory (or table).  Projections at 30TB
would indicate +- 32m files (or records) split on average at 8k per
directory or table.  Personally, I don't think it will be worth our
while to go over the 64TB mark, which means 64m records with 16k
records per table.

In terms of scaling, if we use two physical mysql servers, we're
looking at 2048 tables per server, four servers being 1024 tables, 8
servers being 512 tables per server.  It's relatively easy to
determine from an application point of view, which database server /
database name, and table to query using the same principals that we
are using in terms of storing the files in the associated directory

The issue that we have identified is caused by seek time - hundreds of
clients simultaneously searching for a single file.  The only real way
to explain this is to run 100 concurrent instances of bonnie++ doing
random read/writes... Your disk utilization and disk latency
essentially goes through the roof resulting in IO wait and insanely
high load averages (we've seen it spike to over 150 on a 8-core Xeon -
at which time the application (at a 40 load average already) stops
processing requests to prevent the server crashing).

We are currently busy deploying a small SAN (iSCSI) for testing
changes to the underlying file system, but one part of me believes it
won't help much, whilst the other half is extremely optimistic...
We're also with the SANs splitting the structure so that each SAN only
caters for a certain amount of parent directories.  We're doing 2 SANs
with 2048 directory sets per SAN

On Sat, Jul 27, 2013 at 12:19 AM, Johan De Meersman <vegivamp@stripped> wrote:
> Your argument against FS assumes that you don't know the exact filename (directory
> traversals), but your argument for InnoDB assumes that you do (index lookup). Apples and
> oranges.
> Besides, the venerable ext2 handled up to a couple of tens of thousands of files per
> directory smoothly when listing; things have only improved since then. "Small amounts" is
> a very relative concept.
> Rick James <rjames@stripped> wrote:
>>"Count the disk hits"
>>If you have a filesystem directory, consider that it is designed to
>>handle small numbers of files per directory.  Consider that there is a
>>limited cache for directories, etc.  Plus there is the inode (vnode,
>>whatever) storage for each file.  I don't know the details (and it
>>varies wildly with "filesystem" (ext, xfs, zfs, etc)).
>>Looking at InnoDB...
>>Let's say you have a billion rows in a single table, and you need to
>>fetch one row by the PRIMARY KEY, and it is a MD5 (sha-1, UUID, etc).
>>Such a key is _very_ random.
>>A billion rows would need about 5 levels of BTree.  The top levels
>>would quickly all be cached.  (100M blocks * 16KB = 1.6GB.)  If the
>>leaf nodes add up to 200GB, that is probably bigger than you
>>innodb_buffer_pool_size.  In that case, a _random_ fetch is likely to
>>be a cache miss.
>>A cache miss is about 100ms on normal rotating-media; perhaps 10ms on
>>SSDs.  This limits your reads to 10 (or 100) per second.
>>If you have big BLOBs in the table, then it gets messier.  InnoDB does
>>not put more than 8K of a row in the actual 16KB block.  The rest is
>>stored in another block(s).  So, it is likely to take an extra disk hit
>>If your data size is 100 times as big as your buffer pool, then it
>>becomes likely that the next level of the BTree won't be fully
>>cacheable.  Now 300ms/30ms.
>>I think it is likely that the small number of disk hits for InnoDB is
>>better than the many disk hits for traversing a directory tree (with
>>large directories) in the filesystem.  I vote for InnoDB over the
>>directory tree.
>>Yes, you will have seeks.
>>No, adding more RAM won't help much.  Here's an argument:
>>Suppose your data is 20 times as big as the buffer pool and you are
>>doing random fetches (MD5, etc).  Then 1/20 of fetches are cached; 95%
>>cache miss.  Estimated time: 0.95 * 100ms = 95ms.
>>Now you double your RAM.  1/10 cached -> 90% cache miss -> 90ms average
>>-> Not much improvement over 95.
>>> -----Original Message-----
>>> From: cknipe@stripped [mailto:cknipe@stripped] On Behalf Of
>>> Chris Knipe
>>> Sent: Friday, July 26, 2013 12:30 AM
>>> To: Johan De Meersman
>>> Cc: mysql
>>> Subject: Re: hypothetical question about data storage
>>> Hi All,
>>> Thanks for the responces, and I do concur.  I was taking a stab in
>>> dark so to speak.
>>> We are working with our hosting providers currently and will be
>>> introducing a multitude of small iSCSI SANs to split the storage
>>> structure over a multitude of disks...   This is something that needs
>>> to be addressed from a systems perspective rather than an
>>> one.
>>> SSD (or Fusion and the like) are unfortunately still way to expensive
>>> the capacity that we require (good couple of TBs) - so mechanical
>>disks it
>>> would need to be.  However, with the use of SANs as we hope, we
>>should be
>>> able to go up from 4 to over 64 spindles whilst still being able to
>>> the storage and have redundancy.
>>> Many thanks for the inputs and feedbacks...
>>> --
>>> C
>>> On Fri, Jul 26, 2013 at 9:23 AM, Johan De Meersman
>>> wrote:
>>> > Hey Chris,
>>> >
>>> > I'm afraid that this is not what databases are for, and the first
>>> you'll likely run into is amount of concurrent connections.
>>> >
>>> > This is typically something you should really tackle from a systems
>>> perspective. Seek times are dramatically improved on SSD or similar
>>> storage - think FusionIO cards, but there's also a couple of vendors
>>> (Violin comes to mind) who provide full-blown SSD SANs.
>>> >
>>> > If you prefer staying with spinning disks, you could still improve
>>> seeks by focusing on the inner cylinders and potentially by using
>>> sector formatting. Again, there's SANs that do this for you.
>>> >
>>> > Another minor trick is to turn off access timestamp updates when
>>> mount the filesystem (noatime).
>>> >
>>> > Also benchmark different filesystems, there's major differences
>>> them. I've heard XFS being recommended, but I've never needed to
>>> for seek times myself. We're using IBM's commercial GPFS here, which
>>> good with enormous amounts of huge files (media farm here), not sure
>>> it'd fare with smaller files.
>>> >
>>> > Hope that helps,
>>> > Johan
>>> >
>>> > ----- Original Message -----
>>> >> From: "Chris Knipe" <savage@stripped>
>>> >> To: mysql@stripped
>>> >> Sent: Thursday, 25 July, 2013 11:53:53 PM
>>> >> Subject: hypothetical question about data storage
>>> >>
>>> >> Hi all,
>>> >>
>>> >> We run an VERY io intensive file application service.  Currently,
>>> >> problem is that our disk spindles are being completely killed due
>>> >> insufficient SEEK time on the hard drives (NOT physical read/write
>>> >> speeds).
>>> >>
>>> >> We have an directory structure where the files are stored based on
>>> >> the MD5 checksum of the file name, i.e.
>>> >> /0/00/000/000044533779fce5cf3497f87de1d060
>>> >> The majority of these files, are between 256K and 800K with the
>>> >> exception (say less than 15%) being more than 1M but no more than
>>> >> in size.  The content of the files are pure text (MIME Encoded).
>>> >>
>>> >> We believe that storing these files into an InnoDB table, may
>>> >> actually give us better performance:
>>> >> - There is one large file that is being read/written, instead of
>>> >> BILLIONS of small files
>>> >> - We can split the structure so that each directory (4096 in
>>> >> sit's on their own database
>>> >> - We can move the databases as load increases, which means that we
>>> >> can potentially run 2 physical database servers, each with 2048
>>> >> databases
>>> >> each)
>>> >> - It's easy to move / migrate the data due to mysql and
>>replication -
>>> >> same can be said for redundancy of the data
>>> >>
>>> >> We are more than likely looking at BLOB columns of course, and we
>>> >> need to read/write from the DB in excess of 100mbit/s
>>> >>
>>> >> Would the experts consider something like this as being feasible?
>>> >> it worth it to go down this avenue, or are we just going to run
>>> >> different problems?  If we are facing different problems, what can
>>> >> possibly expect to go wrong here?
>>> >>
>>> >> Many thanks, and I look forward to any input.
>>> >>
>>> >
>>> > --
>>> > Unhappiness is discouraged and will be corrected with kitten
>>> --
>>> Regards,
>>> Chris Knipe
>>> --
>>> MySQL General Mailing List
>>> For list archives:
>>> To unsubscribe:
> --
> Sent from Kaiten Mail. Please excuse my brevity.


Chris Knipe
hypothetical question about data storageChris Knipe25 Jul
  • Re: hypothetical question about data storageVahric Muhtaryan25 Jul
  • Re: hypothetical question about data storageJohan De Meersman26 Jul
    • Re: hypothetical question about data storageChris Knipe26 Jul
      • RE: hypothetical question about data storageRick James26 Jul
        • RE: hypothetical question about data storageJohan De Meersman26 Jul
          • Re: hypothetical question about data storageChris Knipe26 Jul
            • Re: hypothetical question about data storagehsv27 Jul
            • Re: hypothetical question about data storagewilliam drescher27 Jul
              • RE: hypothetical question about data storageRick James29 Jul
                • RE: hypothetical question about data storageJohan De Meersman29 Jul
                  • RE: hypothetical question about data storageRick James29 Jul
                    • Re: hypothetical question about data storageCarsten Pedersen30 Jul
                    • Re: hypothetical question about data storageManuel Arostegui30 Jul
                    • RE: hypothetical question about data storageJohan De Meersman30 Jul