In the film industry we do have many problems that parallelize very well. Physical simulations is one example that people are often familiar with using parallel processing. Rendering is an extremely good example because we can often get parallelism not only within a single image, but because a film is made up of many images, we get additional parallelization by processing multiple images at the same time. My company, Digital Ordnance, builds a high performance image capture and play back system. We take advantage of both CPU and GPU parallelism quite heavily in our systems.

But, the thing that struck me about the post, is that concept that large numbers of cores are applicable to a wide audience. I’m really not convinced that’s true.

My biggest complaint is that for the typical home or business users computers are faster than they need to be. Most users are surfing the web, reading email, running some office applications, and maybe playing some light games. They’re basically doing one task that’s limited in performance by human input. For them pretty much any modern computer has more than enough power. They’d be better suited by having reliable disk storage (RAID), better screens, or cheaper systems, than they would be having faster CPUs. I’m sure there will be some killer app in the future that will set a different baseline for computers (maybe good speach recognition for example), but until that happens they don’t need faster computers.

I think that handles the majority of the computer users out there. But let’s move on to the other case of users who really do need powerful machines.

First, we need to recognize that adding more cores often comes at the cost of slower clock rates. The total number of cycles per a second if you combine all the cores is greater, but the speed of each individual core is slower. That means that to take advantage of all those cores, you have to have a problem that parallelizes across multiple cores. If your software only runs on one CPU you get slower not faster.

Second, we reflect on Ahmdal’s Law. Ahmdal’s law basically says that the speed up you get by parallelizing a problem is limited by the amount of time that must be done serially. For example, if you have a problem where 90% of the computation can be done in parallel, and 10% of the computation is spent managing the CPUs, distributing data, and collating results. As you add more cores the time spent on the parallel part of the computation goes down, but the serial computation remains fixed (or sometimes gets worse, but we’ll talk about that later). As you approach an infinite number of cores, the parallel processing time goes to 0. But that 10% of the time remains constant. That means that if 90% of your compute time can be parallelized, no matter how many cores you throw at it, it will never run more than 10x faster than one core. If your code is 99% parallizable, then you can not get better than 100x faster. That’s the best case, but the situation is actually worse than that.

We already mentioned one of the issues that limits that performance. As you increase the number of cores, the speed of the cores typically decreases. I realize that Ghz isn’t really a good metric, but for this exaxmple, I’m going to use it. I can buy a 3.2Ghz single core processor or I can buy a 2.0 Ghz quad core processor. That quad core gives me 8 Ghz. that should be much faster than 3.2Ghz, right? Well, let’s look at my 90/10 problem above. Let’s say it takes 10 seconds on the one processor. If we ran that on one of our 2.0 GHz cores, it should take 3.2 Ghz/ 2.0 Ghz times longer which is 16 seconds. Now we know that 10% of the time 1.6 seconds is contant, the other 14.4 seconds we’ll assume parallelizes perfectly across the 4 cores for a time of 3.6 seconds. Our total time is therefore 1.6 + 3.6 = 5.2 seconds. We had 2.5 times the computing power but we only got a 1.9 decrease in compute time. Ahmdal’s law and slower cores used up the rest of that performance.

Now I’ll add another problem to the mix. As Ahmdal said, your performance increase is limited by the time spent in the parallel part of the process. As you add more cores, the time spent managing them goes up. Consider that your problem has to be devided among all the CPUs and the results need to be put back together. The more cores you have the more data you need to move around. Sometimes a data structure needs to be updated by one core at a time to avoid each core writing on top of each other. The more cores you have the more time each of them spends waiting for their turn to write the data structure. These problems are hard to describe in detail, but they do contribute to the time spent doing serial processing. So in an ideal world 10% of your compute time would be spent on serial data regardless of how many cores are involved by in reality the percentage of time goes up as the number of cores goes up.

Finally, I’ll describe the lockstep or partitioning problem. In our example 90/10 example I assumed that the 14.4 seconds would be split perfectly among the 4 cores. It turns out that in many cases splitting a problem up perfectly is very difficult. Let’s take a familiar parallel processing problem like rendering an image. If we had four cores we might just split the screen in to quarters and have each core handle one quarter of the screen. In most cases that would work fairly well. But let’s say my image is blank in one corner and has a bunch of mirrors in another corner. In that case, the core handling the blank section will get done quickly. Where the corner with the mirrors takes more work to calculate and takes longer. Since you can’t display the image until all the cores are finished, that means you end up waiting for the slowest core. You might argue that you could break the image in to more pieces, to make the partitioning more fair, and you’d be right, that does help. You might also devise a clever way that when one core finishes early it takes some of the work from a core that is running more slowly. Again, a great idea that will help. But neither of these solutions solves the problem. These extra steps add overhead which slows down the general case, and in the end you still have some cores finishing before others. The time spent with idle cores lowers our overall performace improvement.

One other problem, that’s tangentially related to these discussion, is that how you parallelize your code is often dependent on the target system. If I know my code is going to run on a machine that has 1000 cores, then I’m going to break my problem down in to very small pieces. In doing so, I add a lot of overhead. If I use that same technique on a system with only four cores, then that extra overhead is much more expensive. Applications do their best to partition their problem in to the right size for the number of cores they have, but that isn’t easy and isn’t always perfect. Parallelizing your application for one platform is hard enough, and parallizing it for a wide range of platforms is much more difficult.

Even with all the nay saying there are some applications or portions of applications will be able to use more cores effectively. There are exciting projects that we’re undertaking where we use as many cores as we can get. But there are also bottlenecks where you want one core to run as fast as possible. It would be interesting to work with an architecture where you had a few really fast cores and many slower cores. The exact balance would vary by application, but it would provide a unique way to handle the serial and parallel programming aspects. That’s a problem for the hardware architects to chew on.

I also believe that specialized cores (such as GPUs) are interesting solutions. They work very well for our application, because we’re doing a lot of image processing, but perhaps specialized cores that performed matrix operations quickly might be useful to other clients. The downside to these non-standard cores is that programming them to take advantage of their power is often difficult. The tools need to be put in place that make it just as easy to use these special cores as it is to use a standard core.

Finally, compilers need to get a lot smarter. They need to determine when there is parallelism to exploit and to automatically generate the code to do it. To do that, you need to be able to characterize these cores well enough that a compiler can decide whether the performance improvement would be more than the overhead. Again if we consider that the you might write the application for a wide variety of platforms, that would require that this sort of optimization be done at run-time. How much just in time compiling would need to be done?

All these cores in a variety of designs is making for a very interesting computing landscape for the years to come. I hope we continue to make the most of it.

ShareThis

Worse than that, I generally don’t want to share all my social activities with everyone I’ve ever met, even on one social network. Maybe that makes me weird, but I don’t tink clients I do business with, need to see pictures from a party I’ve thrown.Maybe only my family should see pictures of my father in the hospital. No social network I’ve joined (and has enough of my connections to be useful) let’s me control what type of contact the person is and easily give them different access rights.

The reality is that I have a real collection of contacts on Linked-In. I’ve found it useful for keeping track of where people I know are currently working and it gives me their current email address so I can reach them. I don’t tend to long on to Linked-In very regularly. Basically I go there when I’m looking up a contact, when someone I know sends me an invitation, and every once in a while to look for new contacts I might have missed. Of course, it’s easier to let them come to me, than it is to chase them down.

All the interest in Facebook got me to join that as well. Now I get a duplicate set of friend requests. Then I also get a ton of requests to be a vampire, ninja, warewolf, race car, etc. Those are big time wasters. All the fun boards etc seem to show is silly content.  My front page has gotten so huge with everyone’s updates that it isn’t really managable. I read today that Facebook is coming up with a new interface, but I think that misses the point. I’m not sure organizing the junk is really worth it.

Rather than just bitching, what do I want? Well I think linked-in basically works. It stays out of my way. It has useful information I can reference. It isn’t too hard to deal with the updates. I think we need a good single sign on system so I don’t have to register with every service out there. Instead I use my single sign on, and your service sets up my account automatically the first time I try to use it. Single sign on also lets me create one friend network. Now when I connect to a new service it knows who my friends are and connects them up autoatically according to a user preference. With those changes, everything gets simpler and more useful for the user. Then my final request would be for services to do a small number of things really well, instead of trying to be the wrapper around everything. Then I can decide what features I want to use, and all my friends can see my content without having to join. Personally, I’d probably stick to a few things like a blog and photo/video sharing, but that’s my choice. if you want to play all the Facebook games, then you can join that too.

I think that would make the world of social networking much easier to deal with, and more flexible. It does go against the money flow. Getting everyone locked in to one service that all your friends are on, is really about extracting the dollars. My suggestion would spread the cash around, so developers won’t like that as much.

ShareThis

In large companies there’s a lot of bureaucracy. There are more people who have to be involved in any project and it gets more difficult to synchronize their schedules. There are more people who have to approve a project, so it takes time to process everything.

So it should be better with a small company, right? Not really. Now the problem is that lack of resources, sometimes that’s money, most often that’s time. In a small company everyone is working so hard that there’s not enough slack to add an additional project.

Either way things take longer than you’d expect.  Maybe there is some ideal middle ground, but I don’t think I’ve ever seen it.

ShareThis

This is the trailer for a reality program called Tiger Team which is going to be airing on CourtTV (which, by the way, will be renaming themselves TruTV in January).

It’s a reality show where they follow a company that does security penetration testing. You contract with them to break in to your facility. They give you a 30 day window in which they are going to attempt the break in, but they don’t tell you when and the don’t want you to warn anyone. During the break in they take/create some proof that they performed the task, and afterward they report back on what they did and how you could improve your security.

Having done some penetration testing in the past, security is still a hobby of mine. I found the show quite interesting. The first episode dealt with breaking in to a high end car dealership and the second episode dealt with breaking in to a high end jeweler. The approaches they took in the two episodes were very reasonable. In fact, they handled them pretty exactly as I would have, but they are much more practiced and have at least one trick I didn’t know about. The level of detail on the show was actually very good. You could follow along with that they were doing. They also didn’t sensationalize it to much.

How real is it? We’ll never know how much was set up by the producers, but my opinion is that it was mostly unscripted. In fact, they described the difficulty of getting the camera man to follow in their path without triggering the the motion detector, and that they had passed on some less safe alternatives for entry.

Someone on another blog mentioned that they were surprised they didn’t include a “kids don’t try this at home” message. I’ll say that here. Kids don’t try this at home. These guys are professionals. If these guys get caught by the police I’m sure they can pull out a letter from the owner saying that it was authorized. If you try this on your own, you’ll like get thrown in jail, hurt, or killed.

My final thoughts from the show:

• Security is only as good as the weakest link. So pay attention to ways people can go around your security.
• Security is inconvenient, more security is more inconvenient. The more security you add the better protected you are, but if you go too hard it becomes so inconvenient you can’t get your job down. (And people will work around that defeating your security.
• There is no such thing as totally secure. (follow on to the previous item)
• It’s a good thing criminals are generally dumb. The techniques used in this show really aren’t that hard. Hopefully criminals won’t get better at doing them by watching it.

ShareThis

We got our groove on Saturday night at Harvelle’s in Redondo Beach with grammy award winning blue artist John Lee Hooker Jr. This was the end of the promotional tour for his latest albumn, Cold as Ice. He puts on a great high energy show. He did a mix a classic blues, some boogaloo, and some funk. He had a couple tribute pieces to his father, but he really stands on his own.

This was my first time to Harvelle’s in Redondo. I’d been to the one in Santa Monica several times. Harvelles Redondo has got to be four times bigger than Santa Monica, but it still has an intimate feel. There are plenty of booths and tables where you can sit down, plus a good sized dance floor. You enter on the ground floor at the pier and then decend a staircase to enter Harvelles.

I was surprised that the place was so empty. Of course, it was the weekend before Christmas, so maybe that explains it. There couldn’t have been 50 people there. So it was just like a private show. Our friend Chris took the photo I linked above. Linda and I got out on the dance floor a couple times. She also got to dance with Chris and a couple other people there.
I had a great time. If you get the chance to see John Lee Hooker Jr, you should take it. It was well worth it. I’m going to keep an eye out for more events at Harvelles at Redondo. It’s got the right vibe.

ShareThis

It’s all about getting the shot done. Before the days of digital effects, they would put Vaseline on the lens of the optical printer to hide the wire. It would blur the print.

Digital effects started in the early 80’s. Most effects were still practical, because the digital effects were prohibitively expensive.

I was involved with films such as Titanic and Fifth Element in the late 90’s. We had a full model shop and a large digital effects facility. For each shot we would decide if it could be done better, cheaper, faster, digitally or as a model. We primarily used expensive SGI systems for the digital effects. I introduced Linux to the film industry for the back-end (non interactive) portion of the process.

These days the model shops are all but gone. Almost everything is done digitally. The results are better and a lot more flexible. The majority of the work is done on commodity hardware using Linux in the big facilities and Windows/Macs in the smaller ones. The price hasn’t gotten any lower, because everyone wants bigger effects and the flexibility allows people to avoid planning and making decisions.

This brings us around to Blade Runner and your post. Our system was used as part of the restoration process. They scanned the images at 4k (4x the data of normal film work) and used our Frame Thrower to view the work to decide what needed to be done and to perform quality control on the results. 25 years later commodity hardware and digital effects go in to fixing things Ridley Scott skipped by shooting in the rain.

ShareThis

I had a question on my last post about how things have been since I ditched the Fios HD DVR and switched to to the HD Tivo. I thought that was worth another post.

At this point, I’m a couple weeks in with my switch to the HD Tivo. My girlfriend and I are extremely pleased with the results. The Tivo just works. It is quick and easy to schedule. It records everything I ask. It’s an amazing difference.

Unfortunately, I can’t say the Tivo has been totally perfect. We record and watch a lot of TV. There have been two cases since we got the box where the HD Tivo has stuttered during playback. I tried to watch a show and it would stutter and stop. I tried the obvious things to get it to reset and they didn’t work. Finally, I rebooted the Tivo and the program played fine. During both of these times the HD Tivo was recording two other shows. Of course, it’s pretty standard for us to be recording two shows and watching a third during any evening and it works fine, so it’s only been these two cases where this occurred. Sounds like a bug, but it’s been infrequent enough that it doesn’t upset me, but I’m hoping Tivo fixes it so I can say the experience has been perfect.
A few other thoughts on the current Tivo setup:

20 hours of HD isn’t enough. Tivo or Fios DVR, I want a larger disk. I’ve heard that the e-sata ports on the Tivo are active, and with a few tricks you can add an external drive. That would be perfect. As soon as I get a chance, I’ll add a 500GB eSATA drive and it’ll be great.

For me, two tuners isn’t enough. I said we watch a lot of TV! Now I’ve got two HD tuners in the living room, and a standard def Phillips Series One Tivo in the bedroom. That’s got us covered. I’ve also got a MythTV box with an HD receiver that I might set up with an off the air HD antenna, just so I can get HD in the bedroom instead of just standard def.

The handling of the different HD modes and aspect ratios is really nice. I’ve selected a fixed mode for the 16:9 TV with gray bars for 4:3 programs. The Tivo also supports zooming the 4:3 content. That way letter boxed programs can fill my 16:9 TV correctly. My Sony XBR does that correctly, so I don’t really need that feature, but if I had the HD Tivo connected to the cheapo HD LCD in the bedroom, it would be great. That set doesn’t handle zoom correctly.

I haven’t really used the internet access features. The games on the box are just filler. The universal swivel search is sort of neat. It’s sort of interesting to track the relationship between various films and actors and such.

If anyone from Amazon or NBC is listening, why the heck do I have to configure one click purchasing in order to watch free programs? I tried to download one of your free premier episodes, until it told me I had to go to Amazon and configure one click purchasing to do it. I understand doing that for the episodes that cost money, but make the free downloads easy! Then if I like it, I’ll configure the one click purchase. Instead I gave up on the process, and I won’t use Amazon Unboxed.

As I mentioned, Verizon doesn’t have multistream cable cards, so they had to put two cable cards in the box. I’m not sure if they’re treating those as separate rentals at $3.00 a piece, or as I heard from someone else, it’s really just an outlet charge, and therefore only billed once. I guess I’ll know when I get the next bill.

Since I have an original Phillips series one Tivo on a lifetime subscription, I’m using the multi-unit discount for my HD Tivo. That saves me $6/month. That’s a nice benefit. It’s a shame you can’t get both the pay-up-front discount with the multi-tivo discount, but that’s OK.

I got rid of two FIOS HD-DVRs and replaced them with one HD-Tivo, and one stand def box. That’s quite a bit of cost savings. The setup in the bedroom is essentially free, because I use my old Tivo which is paid for, and the standard def FIOS TV box which they include if you sign up for a year of service. That’s lower quality, but it’s an OK trade off for the bedroom TV.

I’m very pleased with how I’ve got things setup. I think it’s clearly a better solution than the FIOS HD DVR. It removed a ton of frustration. No regrets at all.

Show on map

ShareThis

I would set up a recording, I would see the light going on saying that it was recording, but when you came back to the recording later it said the recording was 0 seconds long. If you tried to play it, you’d get a message saying that the DVR failed to record the program. At one point I caught it while it was failing to record a program. I stopped the recording and tried to manually do the record, but it continued to create a 0 second long recording. The second major problem was that the guide data from Verizon was very inaccurate. I discussed that in a previous post, but the amount of work I had to do to make sure shows were recorded as I wanted or to delete ones I didn’t want was a big hassle.

Beyond that the little things became a big irritation. The remote was awful. It’s huge and complicated. Several of the operations are inconsistent as to which buttons you press. For example some times you have to press the exit button to leave an operation and other times you have to press the stop button. I had one of the DVRs connected to an LCD TV using HDMI. Before the IMG software that worked fine at 1080i, afterward I got a bright green line on the side of the image. If I dropped the resolution to 720p it worked fine. But that also appears to be a problem with the DVR software. Another big annoying issue was that it would jump to a different screen when a recording ended. IE, you come to the end of a program (but not the end of the recording) and start to browse your recorded programs to decide what to watch next. If the recording ends at this moment, the screen would change and you’d be forced to deal with the program that just ended. As I said in a previous post, most of these are annoying glitches.

We had a few hiccups during the install. The first issue is that Verizon doesn’t appear to have multistream cable cards. That means I have to have two cable cards to use my dual tuners. I’m also suspicious that they’re going to bill me for two as well. pissed that they are billing me for two cable cards instead of one. It’s only $3/month each, but if that’s true they really don’t have an incentive to get multistream cards since they make twice as much revenue by having two.

The second problem is that the installation took a really long time. The installer did all the steps he was supposed to, but it didn’t get us a picture. Then he went through his tricks for fixing it and we still had no luck. After a lot of swapping cards and resetting things, he finally decided to try and activate the second card. That went smoothly. Then after a bit more delay the first card came up. He didn’t know why. My only thought was that the first card might have been doing a firmware update, which can take a while, but we didn’t see any indication on the user interface if that was the case.

Of course, our HD Tivo is great. The interface is clean and easy. It works fine. We’ve had it connected to an antenna for the last couple days while we waited for our FIOS appointment and everything worked great. When I first connected it to the network, many of the on-line functions didn’t work. We couldn’t log in to Yahoo for example. The next day all of that was fine and some new items showed up like swivel search. I’m guessing the Tivo did a software upgrade over night. The convergence between the internet and the television on the Tivo is very nice. It’s a shame the netflix agreement didn’t work out. If I could download my 3 DVDs to my Tivo and not use the mail, that would be a really nice solution.

I’m looking forward to a nice smooth working DVR, instead of a lot of frustration.

ShareThis

The new IMG itself is very nice. They’ve made huge improvements to the interface. It’s prettier. It requires fewer clicks of the remote to get around. The wait time for the interface is reduced, and very few places leave you wondering if the box received your command. The down side is that the interface is still inconsistent. There are places where you can press stop to leave an operation and other places you need exit. Some times left arrow backs you out of one tab in their interface, and in other places it will return you to video. There are still places thover that take 5 seconds to respond. But it’s a huge improvement over the old guide.

The quality of the guide data hasn’t improved. There are still way to many generic entries. You can’t rely on new/repeat flats. etc.

But the really killer is that HD-DVR software itself has now gone south. The 1080i signal on one of my flat panel displays now has a bright green line down the side. There also seems to be a problem where the DVR fails to record shows in some cases. The entry is created saying a show was recorded, but if you look at the length of the recording it is 0 seconds.

The bottom line is that now I can’t rely on the guide data or the recorder itself. This is completely unacceptable. To add insult to injury, this article the price of the HD-DVR is going up from $12.99 to $15.99. I don’t know if that only applies to new subscribers or will be applied to existing ones, but either way that’s a problem.

Over 3 years, my two HD-DVR’s cost $1151. Two HD-Tivo’s cost $560. Two 3 year subscriptions (one with multi Tivo discount) cost $382. (oops can’t use the multi-year discount with the multi-tivo discount) $549. Renting two cable cards from Verizon $252. Total $1234 $1361 and I own the Tivos, I get multi room viewing, and it actually works. It seems like a much better solution. Most of the cost is up front, but I’m also insured that the rates don’t increase. Since you can’t use the multiyear discount it’s less money up front. That seems like a much better deal.

ShareThis

A VCR can record a show at a particular time and let you watch it later. But my Tivo does more than that. First, I tell my Tivo to record a program, not to record a time slot. There’s a lot of benefits to that. If the show airs multiple times a week it catches them all. If the show moves it’s time slot for a special event, it handles that. If the show airs multiple times it knows which ones I’ve seen and which ones I haven’t. If it airs multiple times and can’t record it the first time, it can catch it the second. All of these are critical features of a DVR.

All of these depend on having good and accurate guide data. Unfortunately, the FIOS guide data doesn’t meet the basic requirements to work with a DVR. With my Tivo it was common for me to tell it to record only the new episodes of a program. My FIOS DVR had the same option, but I found that was a really bad choice. Several of the programs I wanted to record showed up as repeats when they were new. In fact, in one case it was the series debut! So I decided I really can’t use the record only new episodes option.

They give you three options for recording programs. 1) New only 2) New and Reruns 3) New and Reruns with duplicates. Since new only doesn’t work, I fall back to option 2 and record new with reruns. Then we run in to the next problem. If a show airs multiple times during the week, it is marked as a rerun, but the DVR doesn’t recognize that it’s a duplicate. So the show gets recorded again. I have to manually delete the scheduled recording or erase the recording. Since the space on the DVR is limited to begin with, it really is better to remove the scheduled recording if you can catch it.

Another one of the recording options they provide is whether to record the show only at the specified time or any time it airs. In theory the later would be a useful choice. If I want to record a program, but there are higher priority shows that are keeping the DVR busy, then the system could record it when it aired at some unusual time. That would be nice, but it’s practically unusable because of the poor duplicate detection. Many of the shows on the smaller networks are aired 10 times in a week, and the DVR would record all 10 episodes. Since it only holds 20 hours of HD, that’s half the disk space.

Then we move to another guide issue. The guide should not only track new and repeat episodes, but it should tell me what the show is about. This is particularly useful for news programs which have a single name “Dateline”, “20/20″, or “60 Minutes”. Instead they some wisecracker has written the guide entry that is used many of the times the show airs. For example:

60 Minutes — The granddaddy (and the Rolls-Royce) news magazines, it set the standard for all that followed, and has kept on ticking on CBS since Sept. 24th, 1968, with its familiar format of three stories (most of them hard news) and a commentary or two.

Dateline NBC — NBC’s ubiquitous prime-time news magazine, the one that finally succeeded after more than a decade of attempts (and the one that has aired in many time slots on the network).

Thank you guide writer. Knowing the history of the program really doesn’t help me decide if I want to watch that particular episode. Frankly the tone of some of the entries is quite obnoxious. If I worked at Dateline I’d be upset. From the point of view of a user, these guide listings are virtually worthless.

Without good guide data, most importantly the unique episode numbers and the new/repeat flags, the FIOS DVR loses a lot of value. It’s a frustrating experience getting it to record the shows I want.

There is light at the end of this tunnel. Verizon is talking about their new interactive media guide (IMG) becoming available this summer. Since it’s August, I hope that means this month. I’ve looked at their website and the new IMG looks impressive. If it is backed with better data, then there may be hope. If not, the fancy glitz is just going to make me more annoyed. I also have to point out that it is possible that the guide data is perfect and it is actually the DVR itself that is misinterpreting the data. From what I’ve seen, it doesn’t look that way to me, but if that does turn out to be the case I’ll owe Verizon an apology.

The next installment, is on the DVR itself. Another stumbling block to FIOS’s TV service.

ShareThis