Tonight I had dinner with Henry Daboub (an SVC expert from Houston, TX) and some clients, who asked what I would blog about tonight, and I figured it made sense to blog about the SVC.
Hu Yoshida clarifies his position about storage virtualization, including the statement: "As a result they can not provide the availability, scalability, and performance of a DS8300. If they could, there would be no need for a DS8300."
Of course, if humans descended from apes, why are there still apes? Now that we have cars, why are there still trains? But perhaps a better question is: now that there are supercomputers, why are there still mainframe servers?
The issue is the difference between scale-up versus scale-out. Scale-up is making a single box as big and beefy as possible. When the SVC was introduced, the major vendors all had scale-up designs: IBM ESS 800, HDS Lightning, EMC Symmetrix. Like the mainframe, they were for customers that wanted everything in a single monolithic container.
SAN Volume Controller was the result of IBM Research asking the question, if you could put anyone's software (feature and functionality) on anyone's hardware (monolithic scale-up design), what combination would you choose? What if the brains inside today's monolithic systems could be snapped into the another vendor's frame? What if you could run SRDF on an HDS box, or ShadowImage on an IBM box? The surprising response was that most customers would want a single software for consistency, but wanted the option to choose from different vendors hardware, to negotiate the best price of the commodity iron. Based on this feedback, the SVC was born.
The idea was simple, put all the brains in a separate appliance. The appliance would do the non-disruptive migrations, the caching, the striping, and all the copy services. This lets the customer chose then the hardware they want, any mix of FC and ATA disk, from any vendor.
The SVC design was based on IBM's long history in supercomputers. Using the same "scale-out" technology, the power comes not from having it all in one monolithic box, but rather in a design that combines small nodes together. While the cache is not globally shared, the data is shared between node-pairs, and the logical-to-physical mapping is routed around to all nodes in a cluster. Each SVC node talks to each other SVC node through the FCP ports, eliminating the need for additional wiring. For the most part, each node does its own separate work, but when it needs to, they can communicate across, just like nodes in a supercomputer.
Both the SVC and the DS8300 Turbo have better than 99.999 percent availability, based on redundant components designed for no single point of failure (SPOF). IBM has sold thousands of each, and they have been in the field enough time that we can make that claim. There is nothing between scale-up versus scale-out that makes on inherently more available than the other.
Both the SVC and the DS8300 Turbo can scale from as little as a few TB of disk, to hundreds of TB of disk. We have yet to meet a customer that is too big for the SVC. The DS8300 Turbo is able to scale by adding up to four extension frames, but is still considered a single box from a scale-up perspective. From a processor perspective, an 8-node SVC cluster has 16 Intel Xeon processors, and the DS8300 has 8 POWER5+ processors (dual 4-way). The key advantage of scale-out is that you can add capacity to the SVC in smaller increments. Jumping from a DS8100 (dual 2-way) to a DS8300 (dual 4-way) is a big jump.
SVC remains the fastest disk system in the industry, based on both the SPC-1 and SPC-2 benchmarks. The latest model now supports 8GB per node, for a total of 64GB for an 8-node cluster. This can be used for both read and write non-volatile storage. By comparison, DS8300 Turbo has 32GB write non-volatile storage, and up to 256 GB of read-only cache. The SVC is able to do 155,519 IOPS, faster than the 123,030 IOPS for the DS8300, and of course faster than anything from EMC, HDS, HP or Texas Memory Systems. Of course, workloads vary, and there might be some workloads where the 256GB of read-only cache of the monolithic DS8300 is the better choice.
- copy services
Both SVC and DS8300 Turbo offer FlashCopy (point-in-time copy), Metro Mirror (synchronous) and Global Mirror (asynchronous). SVC provides the additional benefit that it can perform a FlashCopy from one frame to another, and the ability to migrate data seemlessly from one box to another.
Interestingly, IBM has seen a resurgence in both mainframe sales, as well as interest in supercomputers. Both have their place, based on the workload characteristics, and so IBM will continue to offer both modular scale-out designs, as well as monolithic scale-up designs, to meet the different needs of the marketplace.
technorati tags: IBM, disk, SAN, Volume, Controller, DS8300, Turbo, Hu Yoshida, FlashCopy, Metro Mirror, Global Mirror, SPC, benchmarks, HDS, HP, EMC, mainframe, supercomputer
Modified by TonyPearson
Back in 1986, when I first started with IBM, my first job was working on a software product called Data Facility Hierarchical Storage Manager (DFHSM). This did "Information Lifecycle Management" (ILM) by moving data sets from one storage tier to another. (The phrase "Information Lifecycle Management" was coined by StorageTek in 1991, and later resurrected by EMC a few years ago. As is often typical, things that appear new to the distributed systems crowd, are often well-established concepts in the mainframe arena).
To help explain DFHSM and its sister product Data Facility Data Set Services (DFDSS), an enterprising sales rep in Los Angeles named C.D. Larsen made a video called "Re-arranging the sock drawer". He explained that sometimes you want the socks you wear the most on the top drawer, and socks that you only wear now and again in lower drawers. DFHSM can re-arrange your sock drawer based on policy-based automation, determining which ones you wear most often, and moving the others down the "hierarchy" accordingly.
To explain DFDSS, he pulled out an entire drawer of socks, and move it to another level. DFDSS was able to do volume-level backups and dumps to tape very quickly, since it did not process individual data sets, but rather the entire volume image as a whole. These two products are now DFSMShsm and DFSMSdss components of the DFSMS element of the z/OS operating system.
Mainframes use an interesting naming convention for its data sets. 44 characters, divided up into qualifiers that are 1-8 characters long, separated by periods. For example:
The first qualifier indicated it belonged to me, that it was for my Project A, that it was a testcase, and specifically TEST1 job control language. Arranging them in this order meant that I could easily find all the data needed for project A, but if I wanted to keep all the testcase data together, I might have put that as the second qualifer instead.
On Linux, UNIX and Windows, most people are more familiar with hierarchical file systems, so the same file might be stored as:
Same concept. You set up a taxonomy of they way you want to organize your data, so that related data can be grouped together and easier to manage. Whereas we used to tell customers that "Qualifiers are your friend", we now tell people "sub-directories are your friend". This is true when organizing the files on your laptop, in your Lotus Notes, and in Second Life.
Since starting Second Life last November, I have picked up all kinds of free things along the way, and now have thousands of objects in my "inventory". Basically, its like keeping things in your pocket, when you want it, you just take it out of your pocket, and *poof* it appears magically on the ground. I was having a hard time finding things in my inventory, so I decided to re-arrange with sub-folders. This is done in-world, and I found it best to do this away from other avatars asking "what are you doing?" which can get quite annoying. Find a remote island or the rooftop of some building when doing "house cleaning".
I've arranged my main folders as follows. These all appear on a single screen, and makes it easy to find exactly what I am looking for.
- Body Parts
- Calling Cards
- Lost and Found
- Photo Album
In Second Life, you can make complete "outfits" which include your body shape, skin, eyes, hair, and clothes. However, saving away many outfits means duplicating a lot of items. Therefore, I separated them out. I keep body shape, skin, eyes and hair in the folder "Body Parts" and all of the clothing items under "Clothing". Under clothing, I separated everything out into the major categories:
I could have a separate folder for "socks", but I keep those in the "shoes" folder.
technorati tags: IBM, DFHSM, DFDSS, DFSMS, z/OS, qualifiers, taxonomy, Second Life, inventory
Well, I'm back from Mexico.
The flight back was uneventful, except for the leg from Houston to Tucson. The lady in the window seat had "overallocated storage" and required a "distance extension" on her safety belt. To accomodate her, her husband and I flipped up the "logical partitions" between the seats, and "compressed" to take up less space to accomodate. Luckily, it was only for two hours.
On the flight to Houston, I was asked what kind of drink I wanted, in Spanish, as the crew were all from Mexico. Here's a quick Spanish lesson:
- this stands for drink in general, and can include liquor and soft drinks
- this stands generically for soft drink. They will often use "Coke" to refer to any cola beverage, regardless of brand.
It is interesting that Spanish language in each country is slightly different. The Mexicans I met with and spoke Spanish to immediately recognized I was from South America, and not from Central America. Likewise, folks in Puerto Rico knew I was from somewhere from South America, and not from Mexico or Central America. In Columbia, Argentina, and even Brazil, my speech is more recognizable as being from Bolivia.
Before IBM got into an OEM agreement with Network Appliance, I used to indicate that EMC and NetApp were the "Coke and Pepsi" of the NAS marketplace. IBM had a presence, but it was in the single digits, whereas these two major players had roughly equal marketshare, just as Coke and Pepsi dominate equally the US marketplace. That analogy doesn't work in other countries, as in some cases the country might be more heavily in favor of one or the other.
On my flight over from Houston to Tucson, however, I was asked what kind of "pop" I wanted. I always say "soda" to refer generically to soft drinks, but realize that others say "pop" instead. Not only can Americans be able to detect what part of the country people are from by accent, but also by the words they use.
Now I see a blog that explores in great detail the issue of Pop vs Soda vs Coke.
So, it looks like I'll need to "retire" my Coke vs. Pepsi analogy, not because their marketshare has changed, but because IBM's parntering with NetApp greatly skews the advantage over EMC.
technorati tags: IBM, Mexico, NAS, OEM, NetApp, EMC, Coke, Pepsi, Bolivia, Pop, Soda
Today, I went looking for reading-glasses. Unfamiliar with my surroundings, I asked several people where I might be able to find and purchase these, and was sent in various directions. My first stop was a bookstore. It would make sense that since many people need reading glasses to read the books, that they would sell them there, but no. The staff didn't know where I could go, but pointed me in the direction of a mall. At the mall, I found a pharmacy. Many pharmacies sell reading glasses, so I stopped in, but no, not this one. The pharmacists suggested the super-store nearby. I walked in to the super-store, and asked the first employee where they keep their reading glasses, and they said the other corner. The other corner was the electronics department. It made sense that they sold CDs and DVDs in the same section as the equipment that plays them, but reading glasses? Skeptical, I went to the pharmacy department, and the young and beautiful lady (everyone is young, thin and beautiful here) had me follow her, and she led me back to the electronics department, whereupon she pointed to a rack of sunglasses. I indicated that I need reading glasses, not sunglasses. She pulled one out, and it was indeed reading glasses, 1.25, just what I was looking for. Others were tinted, so you can read the newspaper out in the sunlight. The pair I chose cost only $97 in the local currency.
After reading the last sentence, you might be thinking I am describing my "avatar" in Second Life, but no, I am talking about my search for reading glasses on the streets of Mexico. I am here this week in meetings with IBM Business Partners and sales reps to discuss IBM's latest System Storage products and offerings.
We used to tell people they should "clothe" servers with storage. IBM offers both, so yes it makes sense to offer both as part of a complete solution. However, when you look through a dictionary definition "to clothe" you learn it is to dress, wrap or cover with clothing, an implication that it is external, and perhaps temporary, easily changed, like switching from sunglasses to reading glasses. In Second Life, objects can be "worn", simply by attaching or detaching them to your "avatar". Sometimes clothing serves a purpose, like reading glasses, provides protection, like raincoats, and other times, more decorative, like"icing on the cake" or "gold plating".
This concept was fine 50 years ago, when we were in a server-centric world, and dumb storage devices were attached to very intelligent servers. Back then, we used the derogatory term "subsystems" to emphasize that storage was just part of the server, not a system of its own.
Today, we live in an information-centric world. The information outlives the media, and the media outlives the servers that access it. It is not unreasonable to attach dozens or hundreds of servers to a single storage system, or collection of storage systems. Over 20 percent of IBM System Storage DS8000 series, for example, are attached to Windows rack-optimized or blade servers. Imagine a refrigerator surrounded by dozens or hundreds of pizza boxes. Storage is no longer a subsystem, but a system on its own right, dressed, wrapped or covered by servers that deliver the right information, to the right people, at the right time.
So perhaps we should reverse it, telling people they should "clothe" their storage with servers!
technorati tags: IBM, disk, storage, servers, Second Life, clothing, DS8000, Turbo, subsystem, system
I didn't really have a theme this week, still recovering from jet-lag from my travels through Japan, Australia, China.
Gary Diskman has an amusing blog entry about a Funny disaster recovery job posting. It is not clear if he is being completely tongue-in-cheek, or a bit cynical. However, it rings true that you get what you measure, and some managers look for easy metrics, even if there are unintended consequences.
Western medicine works this way. Rather than paying your doctor to keep you healthy, you pay him per visit, to get refills on prescriptions, check-ups on medical conditions, surgeries and so on. While Eastern medicine is focused on keeping people healthy, Western medicine profits more from resolving "situations".
I have seen similar situations on the "health" of the data center. In one case, the admins were measured on how quickly they bring back up their web-servers after a crash. They had this process down to a science, because they were measured on how quickly they resolved the situation. I suggested switching from Windows to Linux, a much more reliable operating system for web-serving, and showed examples of web-servers running Linux that have been up for 1000 days or more. Management changed the metrics to "average up-time in days" and magically the re-boots all but disappeared, thanks to Linux, but also thanks in part to shifting the incentive structure. Perhaps some of those earlier situations were "artificially created"?
Back in the 1980s, I was working on a small software project that was about 5000 lines of code. In those days, testers were measured by the number of "successful" testcases that ran without incident. Testcases that uncovered an error were labeled as "failures" to be re-run after the developers fixed the code. When I declared my code ready for test, the test team ran 110 testcases, all successfully, and they were all rewarded for meeting their schedule. I, on the other hand, did not accept these results, met with them and told them I would give them $100 each if they could find a bug in my code in the next week. Nobody writes 5000 lines of code without some error along the way, not even me. (As one author put it, more people have left earth's gravity to orbit the planet than have written perfect code that did not require subsequent review or testing. It's so true. Good software is difficult to write.)
The test team accepted the challenge, and found 6 problems, more than I expected, but at least I felt more confident of the code quality after fixing these. As I suspected, the unintended consequence of counting "successful" testcases was that testers would write the most simple, basic, least-likely-to-challenge-boundaries testcases to ensure they meet their numbers. My experiment was costly to me, but more importantly was a wake-up call for the test management, and they realized they needed to re-evaluate their test procedures, metrics and terminology. This was a long time ago, and I am glad to see that the overall "software engineering" practice has matured much over the past 20 years.
So, my advice is to determine metrics that have the intended consequences you want, while avoiding any negative unintented consequences that might undermine your eventual success. People will quickly figure out how to maximize the results, and if you can align their goals to company goals, then everybody benefits.
Well, I'll be blogging from Mexico next week (yes, it is a business trip!). Enjoy the weekend.
technorati tags: IBM, Business Continuity, Disaster Recovery, Gary Diskman, indentives, metrics, testcases, unintented consequences