Opportunistic would be the best way to describe our hiring style at Terracotta. We always have our eyes open for the super sharp so I figured I would post about what we generally look for and see if anyone out there shows up to make more magic happen for us. I talked a bit about what I think people should look for when hiring in my blog on teams.
If you are:
Passionate - Do you actually love having complex problems to solve. Can you not sleep at night thinking about how best to design, factor, improve software. Are you always striving for improvement and learning.
Knowledgeable - I'm not going to list a bunch of frameworks here. The knowledge we mostly look for is about knowing how to solve tough problems. The one thing that is a must is you must know multi-threaded programming. Having experience building some kind of infrastructure software is a big plus. Whether it's building an App server, a database or jms system isn't so important but knowing how to build things that need to be scalable is a big plus (just to be clear, not using those things, building them). For some of our stuff knowing classloaders cold and byte code manipulation is a big plus as well. Most of all you must love and be good at writing software.
Respect/Teamwork - This is complex stuff and we are not a big company so you need to be able to talk to and work with others. No room for people who are pedantic and or self serving. It has to be about the product and the software for you not the title. This is true no matter what role you are looking at filling.
Judgement - Each person has to have excellent judgment. Must be able to focus on what's important. If you don't know something, that's not a problem but communicate, ask questions don't rat hole and don't hide problems.
Intelligence - Not going to lie. You must be smart. You should be a person who can analyze and design complex algorithms. Debug and solve complex problems. We will ask you to answer algorithm questions and write some code in the interview.
I know that's a pretty tough list but if you think what we do is cool and fit most of that please shoot your resume to careers at terracottatech dot com.
For those who don't already know about Terracotta here is a brief overview:
We are a well funded startup based in San Francisco but with developers all over the world. Our product is open source network attached memory for Java and is probably the most interesting and diverse product one could work on stretching from distributed computing, to byte code manipulation.
Saturday, July 21, 2007
Thursday, July 19, 2007
What were those results again?
This is a small follow up to my blogs about anti-patterns. When trying to debug a complex logic or performance problem one of the most important things one can do is take notes. I had a conversation with someone the other day where the person said, "I'm in a rush, I don't have time to take notes about my runs." To this I replied, you don't have time not too.
So often we are in such a rush to solve a problem that we cut the wrong corners. When performance tuning, and/or tracking something down that requires multiple runs or configurations of your software always always always take notes on each run. They don't have to be super formal but you should write down all the details you can think of. Some examples include, what were my settings, what did the cpu usage/machine stats look like. What problems did I run into. Always keep a date/time stamp on the tests. This will prevent the inevitable rerunning of tests because you forgot the results, or mixing up what you have tried and not tried. It only takes one mistake to use up more time than tons of note taking would require.
So often we are in such a rush to solve a problem that we cut the wrong corners. When performance tuning, and/or tracking something down that requires multiple runs or configurations of your software always always always take notes on each run. They don't have to be super formal but you should write down all the details you can think of. Some examples include, what were my settings, what did the cpu usage/machine stats look like. What problems did I run into. Always keep a date/time stamp on the tests. This will prevent the inevitable rerunning of tests because you forgot the results, or mixing up what you have tried and not tried. It only takes one mistake to use up more time than tons of note taking would require.
Friday, July 13, 2007
More Lies - Distributed Performance Testing Anti-patterns Part 3 of 4
In part 3 out of 4 of this blog, much like parts 1 and 2 I will hit on anti-patterns that allow your performance testing of clustered and or distributed software to lie to you. I'll be following up part 3 of this blog, the last 4 anti-patterns, with a blog about a simple distributed testing framework I have begun. Hopefully enough of you will be interested, try it and maybe even contribute to it.
Anti-pattern 7: In-memory vs. Distributed Performance Comparison
Description
Writing a test that compares the speed of adding objects to a local, in-memory data structure vs. adding objects to a clustered data structure.
Problem
To avoid suspense, I'll tell you the results of that test without running it. Adding things to a local, in-memory data structure stakes virtually no time at all. In-memory object changes happen so fast, they are hard to even measure. However, when you are making changes to a distributed data structure, no matter what, those state changes have to be shipped off to another location. This takes instructions to be executed to make this happen on top of the ones used for the original task. This isn't just slower, it is way slower. The comparison between in-memory object changes and distributed object changes is useless.
Solution
Figure out how much data you are going to be clustering and what the usage patterns of that data becoming clustered will be. Then simulate and time that. Once again, focus on total throughput with acceptable latency.
Anti-pattern 8: Ignore Real-world Cross-node Patterns
Description
Reading and writing the same data in every node.
Problem
Generally speaking, whether reading or writing, it is more expensive to access the same data concurrently across all nodes. Depending on the underlying clustering infrastructure, this can be more or less of a problem. If you are using an “everything everywhere” strategy, the performance hit of random access across all the data on all the nodes is less, but the “everything everywhere” sharing strategy generally does not scale well. Most other strategies perform better when data access is consistently read and or written from the same node
Solution
Write your performance tests in a way that allows you to set a percentage for locality of reference. Is an object accessed on the same node 80%, 90%, or 99% of the time? You should usually have some cross-node chatter, but usually not too much—although you should be as realistic to the problem you are trying to solve as possible.
Anti-pattern 9: Ignore Usage Patterns
Description
The performance test either just creates objects or just reads objects
Problem
In the real world, an application does a certain amount of reading, writing, and updating of shared objects. And those reads, writes, and updates are of certain sizes.
Solution
If your app likely changes only a few fields in a large object graph, then that is what your performance test should do. If your app is 90% read from multiple threads and 10% write from multiple threads than that is what your test should do. Make your test be true to what you need when it comes to data and usage.
Anti-pattern 10: Log Yourself to Death
Description
Last, but far from least, doing extra stuff like writing data out to a log chews up CPU. Logging too much in any performance test can render the test results meaningless.
This anti-pattern generally covers any extra CPU usage on a load-generating client that affects the performance test. In general, if one or more of your nodes is CPU bound in a cluster performance test, you likely have not maxed-out the performance of your cluster. Let me say that again, if you are resource constrained on any node, including your load generating nodes (but not including your server if one exists) then you are probably not maxing out what your cluster as a whole can handle. Investigate further.
Problem
If the individual load-generating nodes—or even the clustered nodes—are resource constrained, it is likely to create a false bottleneck in your test. You are trying to figure out the throughput of the cluster and your cluster nodes are likely busy doing other things like logging.
Solution
First, always have machine monitoring on all nodes in a performance test. Any time one of the nodes or load generators becomes resource constrained make sure you test with an additional node and see if it adds to the scale. If a node is unexpectedly resource constrained, then take a series of thread dumps (java only) and figure out where all the time is going.
Alright, that is the end of my anti-pattern list for now. I could probably come up with a few more but I'll save them for another day. The moral of this section of the blog is to be curious and skeptical with your testing results. Don't just ask what the numbers are. Find out why and you will end up a much happier person.
Anti-pattern 7: In-memory vs. Distributed Performance Comparison
Description
Writing a test that compares the speed of adding objects to a local, in-memory data structure vs. adding objects to a clustered data structure.
Problem
To avoid suspense, I'll tell you the results of that test without running it. Adding things to a local, in-memory data structure stakes virtually no time at all. In-memory object changes happen so fast, they are hard to even measure. However, when you are making changes to a distributed data structure, no matter what, those state changes have to be shipped off to another location. This takes instructions to be executed to make this happen on top of the ones used for the original task. This isn't just slower, it is way slower. The comparison between in-memory object changes and distributed object changes is useless.
Solution
Figure out how much data you are going to be clustering and what the usage patterns of that data becoming clustered will be. Then simulate and time that. Once again, focus on total throughput with acceptable latency.
Anti-pattern 8: Ignore Real-world Cross-node Patterns
Description
Reading and writing the same data in every node.
Problem
Generally speaking, whether reading or writing, it is more expensive to access the same data concurrently across all nodes. Depending on the underlying clustering infrastructure, this can be more or less of a problem. If you are using an “everything everywhere” strategy, the performance hit of random access across all the data on all the nodes is less, but the “everything everywhere” sharing strategy generally does not scale well. Most other strategies perform better when data access is consistently read and or written from the same node
Solution
Write your performance tests in a way that allows you to set a percentage for locality of reference. Is an object accessed on the same node 80%, 90%, or 99% of the time? You should usually have some cross-node chatter, but usually not too much—although you should be as realistic to the problem you are trying to solve as possible.
Anti-pattern 9: Ignore Usage Patterns
Description
The performance test either just creates objects or just reads objects
Problem
In the real world, an application does a certain amount of reading, writing, and updating of shared objects. And those reads, writes, and updates are of certain sizes.
Solution
If your app likely changes only a few fields in a large object graph, then that is what your performance test should do. If your app is 90% read from multiple threads and 10% write from multiple threads than that is what your test should do. Make your test be true to what you need when it comes to data and usage.
Anti-pattern 10: Log Yourself to Death
Description
Last, but far from least, doing extra stuff like writing data out to a log chews up CPU. Logging too much in any performance test can render the test results meaningless.
This anti-pattern generally covers any extra CPU usage on a load-generating client that affects the performance test. In general, if one or more of your nodes is CPU bound in a cluster performance test, you likely have not maxed-out the performance of your cluster. Let me say that again, if you are resource constrained on any node, including your load generating nodes (but not including your server if one exists) then you are probably not maxing out what your cluster as a whole can handle. Investigate further.
Problem
If the individual load-generating nodes—or even the clustered nodes—are resource constrained, it is likely to create a false bottleneck in your test. You are trying to figure out the throughput of the cluster and your cluster nodes are likely busy doing other things like logging.
Solution
First, always have machine monitoring on all nodes in a performance test. Any time one of the nodes or load generators becomes resource constrained make sure you test with an additional node and see if it adds to the scale. If a node is unexpectedly resource constrained, then take a series of thread dumps (java only) and figure out where all the time is going.
Alright, that is the end of my anti-pattern list for now. I could probably come up with a few more but I'll save them for another day. The moral of this section of the blog is to be curious and skeptical with your testing results. Don't just ask what the numbers are. Find out why and you will end up a much happier person.
Monday, July 2, 2007
Distributed Performance Testing Anti-patterns Part 2 of 4
In Part 1 of this 4 part blog I hit upon 3 Anti-Patterns that can make one's performance testing a poor representation of reality. Here I'm covering 3 more and will be following up with the last 4 in a few days. After that I'm going to talk about a simple distributed performance testing framework I'm going to give away to try and help people be more successful with this stuff.
Anti-pattern 4: Fake Data Fake Performance
Description:
Using data in a distributed performance test that looks nothing like your real data.
Problem:
Distributed computing solutions use all kinds of strategies to move data between nodes under the covers. Just representing a size of data to be shared ignores those strategies and in many cases misrepresents the performance of a real system with real data under real load, both positively and negatively. You may be testing specially optimized flattening tricks that make the system look faster than it is; likewise, you may be testing a particular case that doesn’t perform well, but that isn’t representative of the true performance of the system with real data.
Solution:
Make sure you test with object graphs that vary in size, type, and depth in similar ways to the data you plan to use in your application. Don't assume Maps of Strings will behave anything like the way real object data will behave.
Anti-pattern 5: Incoherent Cluster
Description:
Some clustering products are coherent, some are not, and some have both modes. Don't ignore whether you are testing the performance using the mode you really need for your application.
Problem:
While it is quite possible to have a coherent cluster that has the same throughput as an incoherent cluster, it is certainly harder to do. Coherently clustered software frameworks require the provider to do some fancy locking, batching, windowing, and coherent lazy-loading tricks that aren't for the faint of heart (in the internals of the clustering engine, that is, not for the application developer). You can't assume that performance between a coherent and incoherent clustering approach will be the same.
Solution:
Make sure that if what you need is coherently clustered data that you are actually testing that way. Also, if it’s coherence you’re after, it’s a good idea to verify the end-state of a performance test to make sure the system actually is coherent. Sort of post test verify phase.
Anti-pattern 6: The World by a Thread
Description:
Distributed tests that only use one thread per node.
Problem:
For most clustered software, the name of the game is throughput with acceptable latency. Pretty much all distributed computing software does batching and windowing to improve throughput in a multi-threaded environment. Maxing out a single thread will usually not even approach the max throughput of the JVM or the system as a whole in the same way that a single node will not.
Solution:
Make sure your test uses multiple threads for generating load in each JVM. Check to see if you are cpu bound on any node. If you are not cpu bound you might have a concurrency issue or just need to add more threads.
Conclusion:
I have 4 more anti-patterns that I'm going to publish next week. Keeping an eye on the full 10 will help greatly reduce mistakes in clustering and distributed computing. Once again I'll then be following up with a framework to help develop and run useful tests.
Anti-pattern 4: Fake Data Fake Performance
Description:
Using data in a distributed performance test that looks nothing like your real data.
Problem:
Distributed computing solutions use all kinds of strategies to move data between nodes under the covers. Just representing a size of data to be shared ignores those strategies and in many cases misrepresents the performance of a real system with real data under real load, both positively and negatively. You may be testing specially optimized flattening tricks that make the system look faster than it is; likewise, you may be testing a particular case that doesn’t perform well, but that isn’t representative of the true performance of the system with real data.
Solution:
Make sure you test with object graphs that vary in size, type, and depth in similar ways to the data you plan to use in your application. Don't assume Maps of Strings will behave anything like the way real object data will behave.
Anti-pattern 5: Incoherent Cluster
Description:
Some clustering products are coherent, some are not, and some have both modes. Don't ignore whether you are testing the performance using the mode you really need for your application.
Problem:
While it is quite possible to have a coherent cluster that has the same throughput as an incoherent cluster, it is certainly harder to do. Coherently clustered software frameworks require the provider to do some fancy locking, batching, windowing, and coherent lazy-loading tricks that aren't for the faint of heart (in the internals of the clustering engine, that is, not for the application developer). You can't assume that performance between a coherent and incoherent clustering approach will be the same.
Solution:
Make sure that if what you need is coherently clustered data that you are actually testing that way. Also, if it’s coherence you’re after, it’s a good idea to verify the end-state of a performance test to make sure the system actually is coherent. Sort of post test verify phase.
Anti-pattern 6: The World by a Thread
Description:
Distributed tests that only use one thread per node.
Problem:
For most clustered software, the name of the game is throughput with acceptable latency. Pretty much all distributed computing software does batching and windowing to improve throughput in a multi-threaded environment. Maxing out a single thread will usually not even approach the max throughput of the JVM or the system as a whole in the same way that a single node will not.
Solution:
Make sure your test uses multiple threads for generating load in each JVM. Check to see if you are cpu bound on any node. If you are not cpu bound you might have a concurrency issue or just need to add more threads.
Conclusion:
I have 4 more anti-patterns that I'm going to publish next week. Keeping an eye on the full 10 will help greatly reduce mistakes in clustering and distributed computing. Once again I'll then be following up with a framework to help develop and run useful tests.
Wednesday, June 27, 2007
Why Your Distributed Performance Tests Are Lying to You: Anti-Patterns of Distributed Application Testing and Tuning - Part 1
Clustering and distributing Java applications has never been easier than it is today (see Terracotta). As a result, writing good distributed performance tests and tuning those applications is increasingly important. Performance tuning and testing of distributed and/or clustered applications is an important skill and many who do it can use a little help. Over my next few blogs I'm going to cover a series of anti-patterns in this area. I'll be following it up with a simple open distributed testing framework that I hope can help people out (hint, hint, the testing framework itself is distributed to best test distributed apps).
Here are the first 3 anti-patterns...
Anti-pattern 1: Single-Node “Distributed” Testing
Description
Running your “distributed” performance test inside a single JVM.
Problem
Depending on the framework, this can tell you either: 1) nothing, because the clustering framework recognizes it has no partners so optimizes itself out or 2) very little—it might give one an idea of maximum theoretical read/write speed for that framework.
Solution
When trying to evaluate the performance of any kind of clustering or distributed computing software, always use an absolute minimum of 2 nodes (Preferably more).
Anti-pattern 2: Single-Computer “Distributed” Testing
Description
Putting all (or just too many) of the resources for a performance test on one machine.
Problem
This has two problems. First, distributed applications running on the same machine have different latency and networking characteristics than distributed applications on different machines. This can hide various classes of problems around pipeline stalls, batching, and windowing issues.
The second problem is a variation on another anti-pattern I will discuss later around resource contention. By running multiple JVMs on one machine you are now contending for CPU, disk, network, and potentially affecting context switch rate, etc.
Solution
The only real way to test a distributed application is to run it in a truly distributed way: on multiple machines. If you must have multiple nodes/JVMs on one machine, make sure you are running one of the many resource-monitoring tools and make sure you aren't resource constrained (I use iostat/vmstat for simple tests).
Anti-pattern 3: Multi-Node, Load Only One
Description
Testing with multiple nodes but only sending load/work to one of those nodes while leaving the others just hanging out doing little or nothing
Problem
Depending on the distributed computing architecture chosen, the nodes that are not receiving load may be actually doing a lot of work. If that's the case, only loading one of the nodes is giving a false sense of performance. Also, in some cases, data is lazily loaded into nodes so only putting load on one node could be putting you in the same boat as the single-node tester where no actual clustering is happening.
Solution
When testing clustering software, make sure you are throwing load at all nodes.
Be sure to check back soon as the next few anti-patterns will cover the data aspects of distributed performance testing...
Here are the first 3 anti-patterns...
Anti-pattern 1: Single-Node “Distributed” Testing
Description
Running your “distributed” performance test inside a single JVM.
Problem
Depending on the framework, this can tell you either: 1) nothing, because the clustering framework recognizes it has no partners so optimizes itself out or 2) very little—it might give one an idea of maximum theoretical read/write speed for that framework.
Solution
When trying to evaluate the performance of any kind of clustering or distributed computing software, always use an absolute minimum of 2 nodes (Preferably more).
Anti-pattern 2: Single-Computer “Distributed” Testing
Description
Putting all (or just too many) of the resources for a performance test on one machine.
Problem
This has two problems. First, distributed applications running on the same machine have different latency and networking characteristics than distributed applications on different machines. This can hide various classes of problems around pipeline stalls, batching, and windowing issues.
The second problem is a variation on another anti-pattern I will discuss later around resource contention. By running multiple JVMs on one machine you are now contending for CPU, disk, network, and potentially affecting context switch rate, etc.
Solution
The only real way to test a distributed application is to run it in a truly distributed way: on multiple machines. If you must have multiple nodes/JVMs on one machine, make sure you are running one of the many resource-monitoring tools and make sure you aren't resource constrained (I use iostat/vmstat for simple tests).
Anti-pattern 3: Multi-Node, Load Only One
Description
Testing with multiple nodes but only sending load/work to one of those nodes while leaving the others just hanging out doing little or nothing
Problem
Depending on the distributed computing architecture chosen, the nodes that are not receiving load may be actually doing a lot of work. If that's the case, only loading one of the nodes is giving a false sense of performance. Also, in some cases, data is lazily loaded into nodes so only putting load on one node could be putting you in the same boat as the single-node tester where no actual clustering is happening.
Solution
When testing clustering software, make sure you are throwing load at all nodes.
Be sure to check back soon as the next few anti-patterns will cover the data aspects of distributed performance testing...
Thursday, June 14, 2007
Latency v Throughput
Which is the faster way to get your cargo across the United States. A plane or a train? Some might think the answer is obvious. A plane travels 500 mph (or so) and a train does maybe 80 mph. Therefore the plane is faster. Or is it? The question is really a matter of latency vs. throughput.
Imagine you have to move a bunch of coal across the country and deliver it to a coal processor. Now say that on the west coast, the receiver of the coal can process 100 units of coal an hour. You have 1 train that can haul 10,000 units of coal and takes 48 hours to get to its destination. You have 1 plane that can deliver 100 units of coal in 12 hours.
If the most important thing was to have the coal soon, then the plane is faster (lower latency). But, if the most important thing is to have the coal-processing pipeline filled on the west coast over time then train is faster (higher throughput). Every 96 hours they get 10k units of coal with the train (remember there’s only one train and, just like the plane, it must make the return trip to the east coast). That works out to about 100 units an hour which is just what you need. With the plane, every 96 hours you get 800 pounds of coal. Not nearly fast enough.
The above discussion may seem obvious but I have this conversation all the time when talking about Software: what is fast and what is slow. I've had people tell me it's impossible to do 10 thousand transactions per second in Terracotta when persistent because the disk seek time is 10 millis. Well they would be right if you serialize things. But in infrastructure software, the game is throughput with acceptable latency and it turns out 10 thousand transactions per second isn't all that hard. With parallelism, batching, and windowing, the disk isn't even usually the bottleneck.
Anyway, just wanted to get the throughput v latency thing off my chest.
Imagine you have to move a bunch of coal across the country and deliver it to a coal processor. Now say that on the west coast, the receiver of the coal can process 100 units of coal an hour. You have 1 train that can haul 10,000 units of coal and takes 48 hours to get to its destination. You have 1 plane that can deliver 100 units of coal in 12 hours.
If the most important thing was to have the coal soon, then the plane is faster (lower latency). But, if the most important thing is to have the coal-processing pipeline filled on the west coast over time then train is faster (higher throughput). Every 96 hours they get 10k units of coal with the train (remember there’s only one train and, just like the plane, it must make the return trip to the east coast). That works out to about 100 units an hour which is just what you need. With the plane, every 96 hours you get 800 pounds of coal. Not nearly fast enough.
The above discussion may seem obvious but I have this conversation all the time when talking about Software: what is fast and what is slow. I've had people tell me it's impossible to do 10 thousand transactions per second in Terracotta when persistent because the disk seek time is 10 millis. Well they would be right if you serialize things. But in infrastructure software, the game is throughput with acceptable latency and it turns out 10 thousand transactions per second isn't all that hard. With parallelism, batching, and windowing, the disk isn't even usually the bottleneck.
Anyway, just wanted to get the throughput v latency thing off my chest.
Tuesday, June 12, 2007
Now that's fast...
Alright, I promise I'll get back to blogging about Java and Terracotta stuff next time but... I've been reading a lot of negative press about Apple's Safari 3 beta and while some it is fair I haven't seen a lot of talk about the good stuff about it. So before people flame me let me start with:
Yes, I know it has security holes and those need to be fixed
Yes, I know it has some bugs (like it doesn't work with Zimbra for me)
But...
It still has all the features from Safari 2:
Good new stuff in 3.0:
This thing is blindingly fast? I haven't taken actual timings but just from eyeballing it this thing is super fast. It is much faster than what was already fast Safari 2.0. And much much faster than Firefox. I don't have the time to do real benchmarks on this but I would love to see some.
Much improved inner search. How many times have I hit command F, typed some text, seen the window move but not be able to find where the highlighted word is. Safari does a really nice animated bubble highlight that is impossible to miss. Kudos to the Apple guys for simple subtle improvements.
Plus I think it is supposed to be more standards compliant and it has this resize textbox feature which I haven't tried yet. update: Works great but worth noting that it only works for multi-line test fields not the single line variety.
Yes, I know it has security holes and those need to be fixed
Yes, I know it has some bugs (like it doesn't work with Zimbra for me)
But...
It still has all the features from Safari 2:
- reset browser so when your surfing on someone else's computer you can clean up everything you've logged into like e-mail.
- Really good tabbed browsing
- Private browsing (for when you want to pause the caching and recording of what your browsing)
- plus RSS, popup blocking and the other usual suspects.
Good new stuff in 3.0:
This thing is blindingly fast? I haven't taken actual timings but just from eyeballing it this thing is super fast. It is much faster than what was already fast Safari 2.0. And much much faster than Firefox. I don't have the time to do real benchmarks on this but I would love to see some.
Much improved inner search. How many times have I hit command F, typed some text, seen the window move but not be able to find where the highlighted word is. Safari does a really nice animated bubble highlight that is impossible to miss. Kudos to the Apple guys for simple subtle improvements.
Plus I think it is supposed to be more standards compliant and it has this resize textbox feature which I haven't tried yet. update: Works great but worth noting that it only works for multi-line test fields not the single line variety.
update2: Someone pointed out that Safari3 also now has WYSIWYG editor support. Should have mentioned it since I used that when writing the blog :-)
Anyway, don't want to sound like a fanboy boy, and I might be alone, but I actually like Safari 3.0 and think windows users should give it a go to.
Anyway, don't want to sound like a fanboy boy, and I might be alone, but I actually like Safari 3.0 and think windows users should give it a go to.
Subscribe to:
Posts (Atom)