Comments on: Should I buy a Fast SSD or more memory? http://www.mysqlperformanceblog.com/2010/04/08/fast-ssd-or-more-memory/ Percona's MySQL & InnoDB performance and scalability blog Sat, 11 Feb 2012 16:45:54 +0000 hourly 1 http://wordpress.org/?v=3.2.1 By: Alexey Kupershtokh http://www.mysqlperformanceblog.com/2010/04/08/fast-ssd-or-more-memory/comment-page-1/#comment-785848 Alexey Kupershtokh Wed, 08 Dec 2010 17:48:18 +0000 http://www.mysqlperformanceblog.com/?p=2402#comment-785848 Does sysbench oltp provide the whole dataset of the same temperature (no hot/cold data sets)? If yes then I think in the real world the whole picture could be different. Let's for example naively take the Pareto principle into account. 20% of data provide 80% of io. In this case if we have buffer pool only for 20% (4gb), we cover 80% of io, which results in 600tps instead of 100tps. Does sysbench oltp provide the whole dataset of the same temperature (no hot/cold data sets)? If yes then I think in the real world the whole picture could be different. Let’s for example naively take the Pareto principle into account. 20% of data provide 80% of io. In this case if we have buffer pool only for 20% (4gb), we cover 80% of io, which results in 600tps instead of 100tps.

]]> By: Justin Swanhart http://www.mysqlperformanceblog.com/2010/04/08/fast-ssd-or-more-memory/comment-page-1/#comment-750731 Justin Swanhart Tue, 20 Apr 2010 22:01:42 +0000 http://www.mysqlperformanceblog.com/?p=2402#comment-750731 Frank, As long as you have a battery backed write cache the nobarrier option is safe, and in fact recommended. See the XFS FAQ(http://xfs.org/index.php/XFS_FAQ): Q. Should barriers be enabled with storage which has a persistent write cache? Many hardware RAID have a persistent write cache which preserves it across power failure, interface resets, system crashes, etc. Using write barriers in this instance is not recommended and will in fact lower performance. Therefore, it is recommended to turn off the barrier support and mount the filesystem with "nobarrier". But take care about the hard disk write cache, which should be off. Frank,

As long as you have a battery backed write cache the nobarrier option is safe, and in fact recommended.

See the XFS FAQ(http://xfs.org/index.php/XFS_FAQ):
Q. Should barriers be enabled with storage which has a persistent write cache?

Many hardware RAID have a persistent write cache which preserves it across power failure, interface resets, system crashes, etc. Using write barriers in this instance is not recommended and will in fact lower performance. Therefore, it is recommended to turn off the barrier support and mount the filesystem with “nobarrier”. But take care about the hard disk write cache, which should be off.

]]> By: Kristian Köhntopp http://www.mysqlperformanceblog.com/2010/04/08/fast-ssd-or-more-memory/comment-page-1/#comment-750498 Kristian Köhntopp Tue, 20 Apr 2010 08:23:47 +0000 http://www.mysqlperformanceblog.com/?p=2402#comment-750498 @4: That is actually a benchmark that I would like to see or try myself: "How much load in terms of qps can you add to a MySQL master until the slave starts to lag? Now, switching from SBR to RBR, how does this affect size of the binlogs and how does it affect performance?" @4: That is actually a benchmark that I would like to see or try myself: “How much load in terms of qps can you add to a MySQL master until the slave starts to lag? Now, switching from SBR to RBR, how does this affect size of the binlogs and how does it affect performance?”

]]> By: Frank http://www.mysqlperformanceblog.com/2010/04/08/fast-ssd-or-more-memory/comment-page-1/#comment-746006 Frank Sat, 10 Apr 2010 09:00:45 +0000 http://www.mysqlperformanceblog.com/?p=2402#comment-746006 Hi, is on this hardware (RAID 10 with BBU, SSD,...) the setting: XFS Filesystem mounted with nobarrier option reliable for production use? regards Frank Hi,

is on this hardware (RAID 10 with BBU, SSD,…) the setting:
XFS Filesystem mounted with nobarrier option
reliable for production use?

regards Frank

]]> By: Jonathan Levin http://www.mysqlperformanceblog.com/2010/04/08/fast-ssd-or-more-memory/comment-page-1/#comment-745620 Jonathan Levin Fri, 09 Apr 2010 12:05:14 +0000 http://www.mysqlperformanceblog.com/?p=2402#comment-745620 *Correction* When I meant "store all the database in memory" I meant, like the article was refering, to have the memory as large as the database size. *Correction*
When I meant “store all the database in memory” I meant, like the article was refering, to have the memory as large as the database size.

]]> By: Jonathan Levin http://www.mysqlperformanceblog.com/2010/04/08/fast-ssd-or-more-memory/comment-page-1/#comment-745600 Jonathan Levin Fri, 09 Apr 2010 10:12:29 +0000 http://www.mysqlperformanceblog.com/?p=2402#comment-745600 On the product information for the FusionIO cards, they mention (or encourage this as an option) to make the FusionIO card the swap partition for a linux server (I have it in a pdf a sales guy gave me). If that is in fact the case, would you be able to store all the database in memory along with the FusionIO-swap and let the system shift around with the blocks that are more commonly used? On the product information for the FusionIO cards, they mention (or encourage this as an option) to make the FusionIO card the swap partition for a linux server (I have it in a pdf a sales guy gave me).

If that is in fact the case, would you be able to store all the database in memory along with the FusionIO-swap and let the system shift around with the blocks that are more commonly used?

]]> By: Kevin Burton http://www.mysqlperformanceblog.com/2010/04/08/fast-ssd-or-more-memory/comment-page-1/#comment-745526 Kevin Burton Fri, 09 Apr 2010 06:19:55 +0000 http://www.mysqlperformanceblog.com/?p=2402#comment-745526 We just did an exhaustive test of SSD. Long story short, I think the Intel X-25E SSDs *can* replace a MySQL DB but you have to shard it. Instead of one big DB box with 64GB of memory and 6x SAS you'll need to use 6 DB instances each with it's own replication thread to another server and have the data on 6 SSDs backed by a JBOD. The problem is that MySQL replication is single threaded and this basically destroys performance because adding 6 or 100 SSDs still yields the same performance as one SSD. We just did an exhaustive test of SSD.

Long story short, I think the Intel X-25E SSDs *can* replace a MySQL DB but you have to shard it.

Instead of one big DB box with 64GB of memory and 6x SAS you’ll need to use 6 DB instances each with it’s own replication thread to another server and have the data on 6 SSDs backed by a JBOD.

The problem is that MySQL replication is single threaded and this basically destroys performance because adding 6 or 100 SSDs still yields the same performance as one SSD.

]]> By: Patrick Casey http://www.mysqlperformanceblog.com/2010/04/08/fast-ssd-or-more-memory/comment-page-1/#comment-745460 Patrick Casey Fri, 09 Apr 2010 02:54:11 +0000 http://www.mysqlperformanceblog.com/?p=2402#comment-745460 The answer here appears to be "buy more memory" doesn't it? A 20G box on RAID 10 runs rings around a 16G box on Fusion IO, and for the price of one Fusion IO card, I can buy a lot of memory. I can conceive of plenty of cases where I want fast storage, but if my workload looked like this I'd throw memory at the box for years before I considered upgrading the storage system. Even a mid range 2U dell server will take > 128G of memory these days. The answer here appears to be “buy more memory” doesn’t it?

A 20G box on RAID 10 runs rings around a 16G box on Fusion IO, and for the price of one Fusion IO card, I can buy a lot of memory.

I can conceive of plenty of cases where I want fast storage, but if my workload looked like this I’d throw memory at the box for years before I considered upgrading the storage system.

Even a mid range 2U dell server will take > 128G of memory these days.

]]> By: Charles http://www.mysqlperformanceblog.com/2010/04/08/fast-ssd-or-more-memory/comment-page-1/#comment-745335 Charles Thu, 08 Apr 2010 20:26:15 +0000 http://www.mysqlperformanceblog.com/?p=2402#comment-745335 > There are already appliances on the market which use 8 Intel SSD devices Eight? How about 24: http://www.siliconmechanics.com/i27100/storage-server.php Of course, when you put 24 64GB SLC drives in there, it's suddenly $25k, so that might not be super practical. ;) > There are already appliances on the market which use 8 Intel SSD devices

Eight? How about 24: http://www.siliconmechanics.com/i27100/storage-server.php

Of course, when you put 24 64GB SLC drives in there, it’s suddenly $25k, so that might not be super practical. ;)

]]> By: Morgan Tocker http://www.mysqlperformanceblog.com/2010/04/08/fast-ssd-or-more-memory/comment-page-1/#comment-745307 Morgan Tocker Thu, 08 Apr 2010 16:42:53 +0000 http://www.mysqlperformanceblog.com/?p=2402#comment-745307 Vadim, I was wondering what you think about SSDs for some more special edge cases. i.e. In terms of 'operational pains' warming a cache post-restart on a device capable of tens of thousands of IOPS should be better? Also - for a (single threaded) slave you may care more about service time than throughput of the device. Vadim, I was wondering what you think about SSDs for some more special edge cases. i.e.

In terms of ‘operational pains’ warming a cache post-restart on a device capable of tens of thousands of IOPS should be better? Also – for a (single threaded) slave you may care more about service time than throughput of the device.

]]>