After the exhaustive building and testing process, we've found several areas where we could have improved the original build.
Improved CPU
When we initially decided which hardware components to use, we thought we would not need very much CPU. While we are not doing any type of parity with our storage, we neglected to account for the checksumming that ZFS does to maintain data integrity. This checksumming consumes significantly more processor time than we had originally anticipated. Many tests were using 70% or more of the CPU. We believe that at this high of CPU utilization that there is significant IO contention. Our next ZFS based storage system will probably be based on a dual socket platform and higher clocked (possibly more cores also) CPU's, giving additional headroom for the checksumming and allowing you to use more advanced features that consume CPU resources like Deduplication and Compression. It is not a noticeable problem when testing with gigabit Ethernet speeds. We have been doing some additional benchmarking using 20Gbps InfiniBand, and we have been able to max out the CPU in the ZFS server well before we approached the limits of 20Gbps networking.
More Memory
Going into this project, we did not really know how much main memory we would need in the ZFS SAN, or how well the system would perform with more main memory. After doing some tests on smaller datasets that fit entirely into main memory, we decided that our next build would be 48GB of RAM or more. As a general rule, ZFS will benefit from as much RAM as you can afford to give it. The ARC (main memory) cache of Nexenta and OpenSolaris both function great when the dataset fits entirely into the main cache, and the performance benefits gained from having significant amounts of main memory are huge. At some point you will run into diminishing returns. If you're working with a dataset that is able to fit into main memory and is mainly reads, having more memory for the ARC cache will significantly improve performance. We saw numbers in the 100's of thousands of IOPS when working just out of main memory for random reads. On the flip side of the coin, if your workload is mainly writes then adding 48GB of RAM or more may not give you any noticeable performance advantage.
SAS drives
We thought ZFS's advanced software could overcome some of the inherent problems with slow spindle speeds, and it did up to a certain point. ZFS on OpenSolaris was able to outperform the Promise M610i at basically the same price point. However, we feel we left a lot more performance on the table. Next time we deploy a ZFS server, we plan to use 15k RPM SAS drives instead of 7200 RPM SATA drives as the primary storage. We suspect that we could have easily doubled the performance of our ZFS box in certain tests by using 15k RPM SAS drives. The downside of the SAS drives will be increased cost and decreased capacity, but those tradeoffs will be worthwhile for us if we can double the IOPS, especially on write operations where all transactions have to be committed to disk as quickly as possible. Reads may not be affected as much since many of the reads are coming from SSD storage already, and having SAS drives feed the SSD's would probably not increase overall performance unless your working set is large enough to exceed the total capacity of the SSD drives.
SSD Drives
In the ZFS project, we used SLC style SSD drives for ZIL and MLC style SSD drives for L2ARC. If the price on MLC style SSD drives continues to fall, we will eventually omit the L2ARC and simply use MLC style SSD drives for all of the primary storage. When that day comes, we will also need to use multiple SAS controllers and a much faster CPU in each ZFS box to keep up with all of the IO that it will be able to deliver. Our only concern would be the wear leveling on the MLC drives and the ability of the drives to sustain writes over an extended period of time. Only time will tell if the drives will be able to handle the sustained writes in an L2ARC role or as a primary storage role.
If you decide to use MLC SSD drives for actual storage instead of using SATA or SAS hard drives, then you don’t need to use cache drives. Since all of the storage drives would already be ultra fast SSD drives, there would be no performance gained from also running cache drives. You would still need to run SLC SSD drives for ZIL drives, though, as that would reduce wear on the MLC SSD drives that were being used for data storage.
If you plan to attach a lot of SSD drives, remember to use multiple SAS controllers. The SAS controller in the motherboard for our ZFS Build project is based on the LSI 1068e chipset. We could not find specific numbers for our integrated SAS controller, but another LSI 1068 based standalone card the LSI SAS3080X-R is able sustain 140,000 IOPS. If you use enough SSD drives, you could actually saturate the SAS controller. As a general rule of thumb, you may want to have one additional SAS controller for every 24 MLC style SSD drives. Of course, we have not tested with 24 MLC style SSD's, that number could be higher or lower, but based on our initial performance numbers and the percieved performance of our SAS controller, we believe that 24 would be a good starting point.
Facebook
Twitter
RSS
102 Comments
View All Comments
Exelius - Wednesday, October 6, 2010 - link
I think you identified the strong issue between SATA and SAS drives, but there's no real reason you can't do both: in fact, this is common practice. I don't know what the distribution for AT is so I may be wrong, but often a relatively small amount of your data is accountable for a large portion of your random writes. Why not store that data permanently on the SSDs?For everything else, the cost per gb difference between SATA and SAS is too much to ignore. Once you start talking about adding SAS drives to this, you're moving out of the same class as the Promise device. I've used the Promise vTrak M series (and actually, the M610i specifically) and it's about the cheapest iSCSI SAN device you can get while still being a "real" iSCSI device. It's also about at least a 5 year old product and is growing long in the tooth; I don't know that it's appropriate to compare it with a brand new, performance tuned monster.
But once you introduce SAS into the equation, the chassis itself becomes a much smaller percentage of cost. You go from $140 a drive to close to $400. You also start competing with EqualLogic, HP, etc. and given the need you expressed to add more RAM and CPU, there's definitely some stiff competition from higher-end, more modern products than the M610i.
I guess at the end of the day, while the performance numbers are impressive compared to the M610i, I don't know that the M610i is the device I would use if I was interested in performance. The Promise M610i's strength is price and capacity. Given that the M610i is INFINITELY easier to set up and maintain, that has to factor in to the cost as well. The M610i is often used as a staging target for disk-disk-tape backups; it actually has some throughput issues in a number of scenarios so it's not appropriate for all situations. It just depends on where your needs and bottlenecks are.
I'd rather have seen a comparison with a device such as an EqualLogic or StorageWorks array; because once you upgrade the ZFS box, add labor and support costs into the equation, they do become more appealing in the $10k range (and the fact that you can rather easily add more spindles to an existing array.)
Mattbreitbach - Wednesday, October 6, 2010 - link
You make some strong points.1 - our storage system is not used at Anandtech in any way - I am involved in an entirely separate entity who's only affiliation with Anandtech is that we've written an article reviewing our hardware in our environment. As such, I have no idea what Anandtech's storage needs look like. In our environment we use fixed size VHD's for our VM storage currently. As such there is no real way to put small writes on SSD's and static content on slower spindles. We need to maintain performance across the entire data set.
2 - The Vtrak M610i is about 3 years old from what I can gather from their press releases. We purchased our first Vtrak M610i at about that time. http://www.promise.com/news_room/news.aspx?m=615&a...
While it may be getting a bit older, it is still available for purchase, and is still a relatively inexpensive way to build a high-capacity SAN device. The reason that it was compared in this article is because that is what we are currently using and replacing. While the controller and chassis is different from our ZFS monster, the drives in the chassis are identical, and the price points are very similar.
3 - We would have loved to compare it to a current generation Equalogic unit, but we did not have one on hand to test. If we ever happen to get one we will definately run the numbers against it.
4 - The Promise system has a lot going for it in the ease of setup and use department, and I am currently working on an article that goes in depth on that. Promise also has several new products available that lower the price point (VessRAID) and expand the options that you have available. I hope to get one of those units to test and possibly deploy in the near future also. They also have an enterprise-grade head end (Vtrak S3000) that looks promising also.
Overall, this article was mainly about the ZFS system, what is possible, and how it performed against our current infrastructure. I am hopeful that we can expand what we have on hand to test with and provide broader comparisons in the future, but there is only so far a budget will stretch for getting hardware to simply test.
Exelius - Thursday, October 7, 2010 - link
I know it's at least 4 years old -- I purchased one at least that long ago. But point taken; I haven't kept up with Promise beyond the vTrak M after getting a budget to higher-end units (I still used the vTrak Ms for cheap storage.)And if your data set is large enough to require this many spindles, you might benefit from optimizing it a bit on the front-end... for example, build your VMs to split the VHDs so the high-write data is stored elsewhere. No idea if this would be of benefit for your environment (that's what test labs are for) but it's a strategy most shops with high-volume, high-transaction datasets have to periodically look at as the performance gulf between big, cheap drives and small, fast ones keeps increasing.
Given the size of your environment, EqualLogic or StorageWorks would probably be willing to let you use a demo unit for a little while. Don't know that they wouldn't make you sign an NDA regarding the benchmarks, but you'd at least be able to do it internally... Plus, IMO, there's a massive benefit to having an enterprise support contract when you have a controller failure (which I'm actually surprised hasn't been an issue with the single controller design of the Promise M610)
All told; still a good article -- you generally don't see stuff this thorough posted on the Internet. There are just so many possibilities in this space that it's hard not to nitpick. :)
JonBendtsen - Thursday, October 7, 2010 - link
I think it could be interesting to see performance benchmarks without the L2ARC to see how much value it really has.binarycrusader - Thursday, October 7, 2010 - link
Management of the drive LEDs for faulty drives, etc. is available with the right hardware; it's unfortunate that's it's not well supported on a wide variety of systems, but it does exist.As for SMTP notification (and other kinds) of faulty hardware, etc. that should be available depending on the build of OpenSolaris you're using and whether fault management aware drivers are available for your hardware. See 'man fmadm', 'man fmd' and 'man smtp-notify' for more information.
Ultimately, users looking for a polished storage system with graphical management tools, etc. are encouraged to look at Oracle's Sun Open Storage servers which address many of the complaints listed in the article. Yes, I'm aware you're trying to build your own systems here, but it should be obvious why all of the nice tools aren't given away for free.
pburdine - Friday, October 8, 2010 - link
I haven't installed OpenSolaris yet, but when I am using Solaris 10 with ZFS, it does come with a website manager to manage may of the Sun/Oracle Applications. Did you try https://localhost:6789?murdmath - Friday, October 8, 2010 - link
Great Article. Very informative. I am excited for you review of the Promise M610i SAN.Mat
Brutalizer - Monday, October 11, 2010 - link
First of all, there is only ONE single reason to use ZFS: it protects your data whereas other storage solutions might corrupt your data (including enterprise storage solutions)!See here how common file systems such as Ext3, JFS, XFS, ReiserFS, NTFS, etc might corrupt your data:
http://www.zdnet.com/blog/storage/how-microsoft-pu...
All the rest of the ZFS features such as snapshots, easy administration, etc are just icing on the cake. If ZFS had only protection of data and no other features, I would still use ZFS..
See here how Raid-5 does not protect your data. In fact, Raid-6 is not better and also may corrupt your precious data. Google "data corruption raid-6"
http://www.baarf.com/
See here how ZFS does protect your data:
http://www.zdnet.com/blog/storage/zfs-data-integri...
http://queue.acm.org/detail.cfm?id=1317400
There is a reason ZFS eats CPU (does checksumming and protects your data), whereas all the other filesystems does not protect your data (rudimentary checksumming).
ZFS has end-to-end checksumming! That means, ZFS will compare the data in RAM with the data on disk - are they equal? All other storage solutions does not do that - they only check data within a realm. But when data passes a realm it may corrupt (RAM down to disk controller down to disk). There may be bugs in hardware or software within a realm. And data are never compared "the data XYZ in RAM, is it still XYZ on disk?" - this check are never ever made (unless you use ZFS).
Regarding dedup. If you get slow performance of dedup, it is only because dedup requires huge amounts of RAM. You need something like 2GB RAM for each TB disk. If you have less RAM, dedup will be sloooow. If you have much RAM, dedup will be fast.
Another advantage of ZFS (there are many) is that ZFS is OS agnostic! You can insert your zfs raid into another OS or computer without any problems! Try that with a hardware raid - impossible.
Another advantage of ZFS is there are no "fsck"! Instead you do "zfs scrub" every week, while your raid is alive and running. fsck requires you to shutdown the raid to validate it.
Hardware raid is just a cpu with some software running on it. It is better to move that software to the CPU where you have many cores and GB of RAM and you can easily patch it. In the future, hardware raid will die. Software raid like ZFS will rule.
Regarding BTRFS, if you read the mail lists, you see that people loose data all the time with BTRFS. In the future it might be good, but it will take at least another 5 years until we reach that stage. Then ZFS have developed even further.
solori - Friday, October 22, 2010 - link
Regarding ZFS and ubiquity: ZFS is only version compatible. As ZFS' capabilities are updated, the blanket statement that "any ZFS-speaking OS can mount a ZFS volume" just isn't going to ring true. In fact, many distributions porting ZFS are still behind in ZFS version.As in most "backward compatible" entities, newer versions of ZFS will almost always be compatible with older versions, but the older version will not be able to mount a more recent version. Therefore, you could have a Mac port that can't read a BSD port for instance.
Also, since ZFS is modular, one OS vendor could included a "highly proprietary" inline encryption or compression algorithm that is not (or not strictly) open in nature. This leads to subsequent OS-based divergence if they fail to include the necessary libraries that are not a part of ZFS itself.
However, and for the most part, ZFS should be regarded as version compatible regardless of the OS. Another great reason to use JBOD or discrete disk setups: complete portability of storage pools.
Hrel - Monday, October 11, 2010 - link
Why is ZFS not the only file system in use today? I completely forgot about this until this article. I remember first reading about it and thinking "this'll probably be in everything in a couple years" so I put it out of my mind. I am upset this is not the file system everything uses.