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
cdillon - Tuesday, October 5, 2010 - link
I've been working on getting the additional parts necessary to build a similar system out of a slightly used HP DL380 G5 with a bunch of 15K SAS drives and an MSA20 shelf full of 750GB SATA drives. Here's what I'm going to be doing a little differently from what you've done:1) More CPU (already there, it has dual Xeon X5355 if I recall correctly)
2) Two mirrored OCZ Vertex2 EX 50GB drives for the SLOG device (the ZIL write cache). Even though the Vertex2 claims a highly impressive 50,000 random-write IOPS, the ZIL is written sequentially, and the Vertex2 EX claims to sustain 250MB/sec writes, so it should make a very good SLOG device.
3) Two OCZ Vertex2 100G (the cheaper MLC models) for L2ARC.
4) The SSDs will be put on a separate SAS HBA card from the HDDs to prevent I/O starvation due to the HBA I/O queue filling up because of the relatively slow I/O service-times of the HDDs.
5) Quad Gigabit Ethernet or 10G Ethernet link. The latter will require an upgrade to our datacenter switches, which is probably going to happen soon anyway.
mbreitba - Tuesday, October 5, 2010 - link
I would love to see performance results for your setup. The IOMeter ICF file that we have linked to in the article would help you run the exact same tests as we ran if you would be interested in running them.cdillon - Tuesday, October 5, 2010 - link
I forgot to mention it might also be running FreeBSD (which I'm very familiar with) rather than Nexenta or OpenSolaris, but I'm just kind of playing it by ear. I may try all three. The goal is for it to eventually become a production storage server, but I'm going to do a bit of experimentation first. I still haven't gotten around to ordering the SSDs and the extra SAS HBAs, so it'll be a while before I have any benchmarks for you.Maveric007 - Tuesday, October 5, 2010 - link
You should throw Linux into the mix. You find your performance will increase over the other selections ;)MGSsancho - Tuesday, October 5, 2010 - link
ZFS on linux is terrible. also ZFS on FreeBSD is decent. recent ZFS features such as deduplication and iSCSI are not available on FreeBSD. just grab a copy of the latest build of opensolaris (134), compile it from build 157. use solaris 10 (got to pay now), or use one of the mentioned Nexenta distros.From personal experience, use fast SSD drives. I made the mistake of using a pair of the Intel 40GB Value drives for a home box with 8 x 1.5 TB drives. terrible performance. Yes it is cool for latency but I cant get more than 40MBs from it. I have tried using them just for ZIL or just for L2ARC and performance is abyssal. Get the fastest possible drives you can afford.
Matt, have you tested with using for example realtek nics (dont, pain in the ass), intel desktop nics (stable) or the more fancy server grade nics that have reported iSCSI offload? also have you tried using dedup/compresion for increased performance/space savings? this will use up lots of memory for indexies but if your cpus are fast enough along with network, less IO hits the discs. I hear it has worked assuming you have the memory, CPU, network. One last bit, try using the Sun 40GBs infiniband cards? I know they will work with solaris 10 and opensolaris and thus I would assume nexenta. might want to check the hardware compatibility list for your IB card.
Cheers
Mattbreitbach - Tuesday, October 5, 2010 - link
We have not tested with any other NIC's other than the Intel GB nics onboard the blade. We considered using an iSCSI offload NIC for the ZFS system, but given the cost of such cards we could not justify using them.As for Deduplication - we have recently tested using deduplication on Nexenta and the results were abysmal. Most tests were reading above 90% CPU utilization while delivering far lower IOPS. I believe that deduplication could help performance, but only if you have an insane amount of CPU available. With the checksumming and deduplication running our 5504 was simply not able to keep up. By increasing the core count, adding a second processor, and increasing the clock speed, it may be able to keep up, but after you spend that much additional capital on CPU's and better motherboards, you could increase your spindle count, switch to SAS drives, or simply add another storage unit for marginally more money.
MGSsancho - Tuesday, October 5, 2010 - link
from my personal experience i could not agree more for the deduplication. 33% on each core on my phenom 2 for a home setup is insane. Some things like exchange server, it is best to let the application decide what is should be cached but duplication realy make sense for a tier three storage or nightly backup or maybe for a small dev box. Also the drives them selves mater, you want to use the ones that are geared for raid setups. it allows the system to better communicate with it. I wont name a particular vendor but the current 'green' 5400 rpm 2TB drives are terrible for zfs http://pastebin.com/aS9Zbfeg (not my setup) that is a nightly backup array used at a webhosting facility. sure they have great throughput but all those errors after a few hours.andersenep - Tuesday, October 5, 2010 - link
I use WD green drives in my home OpenSolaris NAS. I have 2 raidz vdevs of 4 drives each (initially I used mirrors, but wanted more space). I can serve 720p content to two laptops and my Xstreamer simultaneously without a hiccup...I guess it depends on your needs, but for a home media server, I have absolutely no complaints with the 'green' drives. Weekly scrubs for 1 yr plus with no issues. I did have to replace a scorpio on my mirrored rpool after 6 months. I am quite happy with my setup.solori - Wednesday, October 20, 2010 - link
As a Nexenta partner, we see these issues all the time. Deduplication is not an apples-apples feature. The system build-out and deduplication set (affecting DDT size) are both unique factors.With ZFS' deduplication, RAM/ARC and L2ARC become critical components for performance. Deduplication tables that spill to disk (will not fit into memory) will cause serious performance issues. Likewise, the deduplication hash function and verify options will impact perfomance.
For each application, doing the math on spindle count (power, cost, space, etc.) versus effective deduplication is always best. Note that deduplication does not need to be enabled pool-wide, and that - like in compression where it is wasteful to compress pre-compressed data - data with low deduplication rates should not be allowed to dominate a deduplication-enabled pool/folder.
Deduplication of 15K, primary storage seems contradictory, but that type of storage has the highest $/TB factor and spindle count for any given capacity target. By allocating deduplication to targets folders/zvol, performance and capacity can be optimized for most use cases. Obviously, data sets that are write-heavy and sensitive to storage latency are not good candidates for deduplication or inline compression.
If you do the math, the cost of SSD augmentation of 7200 RPM SAS pools is very competitive against similar capacity 15K pools. The benefits to SSD augmentation (i.e. L2ARC and ZIL->SLOG where synchronous writes dominate performance profiles) is in higher IOP potential for random IO workloads (where the 7200 disks suffer most). In fact, contrasting 600GB SAS 15K to 2TB SAS 7200, you approach an economic factor where 7200 RPM disks favor mirror groups over 15K raidz groups - again, given the same capacity goals.
The real beauty of ZFS storage - whether it be Opensolaris/Illumos or Nexenta/Stor/Core - is that mixing 15K and 7200 RPM pools within the same system is very easy/effective to do. With the proper SAS controllers and JBOD/RBOD combinations, you can limit 15K applications to a small working set and commit bulk resources to augmented 7200 RPM spindles in robust raidz2 groups (i.e. watch your MTTDL versus raidz).
It is important to note that ZFS was not designed with the "home user" in mind. It can be very memory and CPU/thread hungry and easily out-strip a typical hobbyist's setup. A proper enterprise setup will include 2P quad core and RAM stores suited to the target workload. Since ZFS was designed for robust threading, the more "hardware" threads it has at its disposal, the more efficient it is. While snapshots are "free" in ZFS (i.e. copy-on-write nature of ZFS means writes are the same with or without snapshots) but data integrity (checksums) and compression/deduplication are not.
Mattbreitbach - Wednesday, October 20, 2010 - link
Excellent comments! Thank you for your input.As you noted, we found deduplication to be beyond the reaches of our system. With proper tuning and component selection, I think it could be used very well (and have talked to several people who have had very good experiences with it). For the average home user it's probably beyond the scope of what they would want to use for their storage.