Slashdot Mirror
Are RAID Controllers the Next Data Center Bottleneck?
storagedude writes "This article suggests that most RAID controllers are completely unprepared for solid state drives and parallel file systems, all but guaranteeing another I/O bottleneck in data centers and another round of fixes and upgrades. What's more, some unnamed RAID vendors don't seem to even want to hear about the problem. Quoting: 'Common wisdom has held until now that I/O is random. This may have been true for many applications and file system allocation methodologies in the recent past, but with new file system allocation methods, pNFS and most importantly SSDs, the world as we know it is changing fast. RAID storage vendors who say that IOPS are all that matters for their controllers will be wrong within the next 18 months, if they aren't already.'"
Hardware RAID's are not exactly hopping off the shelf and I think many shops are happy with fiberchannel.
Let's do another reality check: this is enterprise class hardware. Are you telling me you can get SSD RAID/SAN in a COTS package that is cost approximate to whatever is available now? Didn't think so....
Let's face it, in this class of hardware things move much more slowly.
http://www.maxineudall.com/2010/02/should-economists-be-sued-for-malpractice.html
That's fine for some things but I really don't want my confidential client work-product mirrored around the world. Despite all the cloud hype there is still a subset of data that I really do NOT want to let outside my corporate walls.
-B-
FTA Since a disk sector is 512 bytes, requests would translate to 26.9 MB/sec if 55,000 IOPS were done with this size. On the other end of testing for small block random is 8192 byte I/O requests, which are likely the largest request sizes that are considered small block I/O, which translates into 429.7 MB/sec with 55,000 requests
I'm not going to believe an article that assumes that because you can do 55K IOPS for 512Byte reads, you can do the same number of IOPS for 8K reads which are 16X larger and then just extrapolate from there. Especially since most SSD's (at least SATA ones) right now top out around 200MB/s and the SATA interface tops out at 300MB/s. Besides there are already real world articles out there where guys with simple RAID0 SSD's are getting 500-600 MB with 3-4 drives using Motherboard RAID much less dedicated harware RAID.
Storage has been the performance bottleneck for so long, it's a happy problem if you actually must increase the bus speeds/cpu processors/get faster memory on raid cards to keep up. Seems to me the article(or at least the summary) was written by someone hadn't been following enterprise storage for very long...
This is correct, there are laws on the books in most countries that prohibit the exposure of medical and other data
to risk by putting it out in the open. Some have even moved to private virtual circuits, and the SAN's with fast
access via solid state storage of active files works fine, and it moves less accessed data to drive storage,
but none the less quite fast and SAS technology is faster than SCSI tech in throughput.
google "32 trillion offshore needs IRS attention"
An example of SAS throughput pushing out 6 Gbps.
http://www.pmc-sierra.com/sas6g/performance.php
google "32 trillion offshore needs IRS attention"
The second question is, with processors coming with 8 cores, why have some separate specialized controller that handles RAID and not just do it in software?
I much prefer s/ware raid (Linux kernel dm_mirror), it removes a complicated piece of h/ware which is just another thing to go wrong. It also means that you can see the real disks that make up the mirror and so monitor it with the smart tools.
OK: if you do raid5 rather than mirroring (raid1) you might want a h/ware card to offload the work to, but for many systems a few terabyte disks are big and cheap enough to just mirror.
There may need to be some minor rethinking of controller throughput for read applications on smaller data sets for SSD. But right now, I regularly saturate the controller or bus when running sequential RW tests against a large number of physical drives in a RAID{1}0 array, so it's not like that's anything new. Using SSD just makes it more likely that will happen even on random workloads.
There are two major problems with this analysis though. The first is that it presumes SSD will be large enough for the sorts of workloads people with RAID controllers encounter. While there are certainly people using such controllers to accelerate small data sets, you'll find just as many people who are using RAID to handle large amounts of data. Right now, if you've got terabytes of stuff, it's just not practical to use SSD yet. For example, I do database work for living, and the only place we're using SSD right now is for holding indexes. None of the data can fit, and the data growth volume is such that I don't even expect SSDs to ever catch up--hard drives are just keeping up with the pace of data growth.
The second problem is that SSDs rely on volatile write caches in order to achieve their stated write performance, which is just plain not acceptable for enterprise applications where honoring fsync is important, like all database ones. You end up with disk corruption if there's a crash, and as you can see in that article once everything was switched to only relying on non-volatile cache the performance of the SSD wasn't that much better than the RAID 10 system under test. The write IOPS claims of Intel's SSD products are garbage if you care about honoring write guarantees, which means it's not that hard to keep with them after all on the write side in a serious application.
1) Most high-end RAID controllers aren't used for file serving. They are used to serve databases. Changes in filesystem technology don't affect them one bit, as most of the storage allocation decisions are made by the database.
2) Assuming that a SSD controller that can pump 55k IOPS w/ 512B I/O's can do the same w/ 4K I/O's is stupid and probably wrong. That is Cringely math; could this guy possibly be as lame?
3) The databases high-end RAID arrays get mostly used for do not now, and never have, used much bandwidth. They aren't going to magically do so just because the underlying disks (which the front-end server never even sees) can now handle more IOPS.
All SSD's do is flip the Capacity/IOPS equation on the back end. Before, you ran out of drive IOPS before ran out of capacity. Now, you get to run out of capacity before you run out of IOPS on the drive side.
Even if you have sufficient capacity (due to the rapid increase in SSD capacity), you are still going to run out of IOPS capacity on the RAID controller before you run out of IOPS or bandwidth on the drives. The RAID controller still has a lot of work to do with each I/O, and that isn't going to change just because the back-end drives are now more capable.
SirWired
SAS technology is faster than SCSI tech in throughput
"SCSI" does not mean "parallel cable"!
Sorry, pet peev, but obviously Serial Attached SCSI (SAS) is SCSI. All Fibre Channel storage speaks SCSI (the command set) all USB storage too. And iSCSI? Take a wild guess. Solid state hard drives that plug directly into PCIe slots with no other data bus? Still SCSI command set. Fast SATA drives? The high end ones often have a SATA-to-SCSI bridge chip in front of SCSI internals (and SAS can use SATA cabling anyhow these days).
Pardon me, I'll just be over here grumbling about this.
Socialism: a lie told by totalitarians and believed by fools.
Of course. NFS provides an easy to use concurrent shared filesystem that doesn't require any cluster overhead or complication like GFS or GPFS.
I think we need a mod option to mod down the article summary: -1, stupid editor.
You had your chance.
For justice, we must go to Don Corleone
Well, ZFS is great, but don't get that mixed up with software RAID. It's not. The storage redundancy algorithms used by ZFS are not the RAID algorithms, such that using ZFS is much better than using EITHER hardware or software RAID.
ZFS provides performance and data integrity assurance that standard RAID does not. Primarily, because filesystem level data is checksummed, and it should be almost impossible for silent data corruption to occur at the storage device level, except cases where the data written actually matches the checksums, (a later 'zpool scrub' should detect it, if ZFS is implemented properly).
But aside from ZFS, software RAID (and even fakeraid/hostraid hardware adapters that perform RAID in the driver) really really suck both in terms of reliability, data integrity, and performance when you need to push things to the maximum, compared to a good hardware RAID controller; software RAID is measurably slower on the same CPU and memory.
SMART provides so little of what you need to be doing to keep a reliable array, it isn't even funny.
Good hardware controllers keep metadata and do frequent consistency checks / "scrubs" / surface scans, to ensure every bit of data is periodically read from every drive, so HDD firmware has an opportunity to fix errors before they become "unrecoverable read errors".
Hardware controllers will also detect when a hard drive is having a problem that cannot be easily identified by software. Hard drives are direcly plugged into the controller; it can detect things such as abnormal command response latencies.
A software controller can't be sure the abnormal latency isn't due to other workload on the bus, or "not a drive failing", so the HW controller is more responsive to failure.
HW contollers also provide writethrough caching, and sometimes have a BBU with a full writeback cache, which drastically helps performance, and reduces the RAID performance penalty, which software RAID doesn't mitigate, but in fact makes worse.
Oh yes, and Good controllers also have monitoring and administration tools for various OSes, including Linux, Windows, and Solaris, produced by the manufacturer.
Many of the good controllers come equipped with audible alarms and terminals for you to plug drive failure LEDs into, so that anyone near the server can know a drive has failed, and which one.
For my own personal data, I'd consider that adequate. For data I'm legally required to keep secret - absolutely not. Your physical security design should force an attacker to steal both your keys and your data, each from a seperate physical location, so that you can destroy one as soon at the other is stolen to prevent data loss. Electronic security of course focuses on compartmentalization and auditing, so that an inside attacker can only steal a small portionof the data, and can be caught an jailed afterwards. That's all pretty basic design.
Also, 256-bit symmetric encryption really is enough - it's firmly beyond the realm of what can be brute-forced, unless some fundamental understanding of physics is wrong. 256-bit AES is only vulnerable to weaknesses in the algorithim being discovered at some future point. If you're paranoid, you're far better off using 2 unrelated 256-bit symmetric algorithms than a symmetric key larger than 256 bits.
Socialism: a lie told by totalitarians and believed by fools.
All the important operations tend to be random. For a file server, you may have twenty people accessing files simultaneously. Or a hundred, or a thousand. For a webserver, it'll be hitting dozens or hundreds of static pages and, if you have database backend, that's almost entirely random as well.
For people consolidating physical servers to virtual servers, you now have two, three, ten or twenty VMs running on one machine. If every one of those VMs tries to do a "sequential" IO, it gets interlaced by the hypervisor into all the other sequential IOs. No hypervisor would dare tell all the other VMs to sit back and wait so that every IO is sequential. That delay could be seconds or minutes or hours.
Now imagine all that, and take into account that the latest Intel SSD gets around 6600 IOPS read and write. A good, fast hard drive gets 200. So you could put thirty three hard drives in RAID 0 and have the same number of IOPS, and your latency would still be worse. All the RAID0 really does for you is give you a nice big queue pipeline, like in a CPU. Your IO doesn't really get done faster, but you can have many more running simultaneously.
Given that SSDs are easily three to four times faster on sequential IO and an order of magnitude faster on random IO, I don't think it's that implausible to believe that the industry isn't ready.
When you calculate IOPS, a good portion small of reads and writes get executed at random places on the disks. When you you make one filesystem write on a raid0 set (depending on how smart the raid0 controller is), it will be locking up several or ALL the disk spindles for that individual read/write.
Actually, that's incorrect. Here's why:
When you make a RAID0 array, you stripe large blocks between all the disks, usually 64K-256K large. If your operation does not cross the block boundary, you only access a single drive. Assuming those random small files are evenly distributed, your IOPS scale almost linearly with drive count.
Does anyone actually still use NFS?
Of course. It's nearly always fast enough, trivially simple to setup, and doesn't need complicated and fragile clustering software so that multiple systems can access the same disk space.