Micro jitter, busy waiting and binding CPUs

By

Performance profiling a new machine

When I work on a new machine, I like to get an understanding of it's limitations.  In this post I am looking at the jitter on the machine and the impact of busy waiting for a new PC I built this weekend. The specs for the machine are interesting but not the purpose of the post.  Never the less they are:

  • i7-3970X six core running at 4.5 GHz (with HT turned on)
  • 32 GB of PC-1600 memory
  • An OCZ RevoDrive 3, PCI SSD (actual write bandwidth of 600 MB/s)
  • Ubuntu 13.04
Note: the OCZ RevoDrive is not officially supported on Linux, but is much cheaper than their models which are.

Tests for jitter

My micro jitter sampler looks at interrupts to a running thread.  It is similar to jHiccup but instead of measuring how delayed a thread is in waking up, it measures how delays a thread gets once it has started running.  Surprisingly how you run your threads impacts the sort of delays it will see once it wakes up.


This chart is a bit dense.  It shows the number of interrupts of a range of time occurred within a CPU hour on average (each test was run for more than two clock hours)  There raw data is available HERE

The interesting differences are in how binding to an isolated CPU and/or busy waiting threads are handled by the OS.

Busy waiting

In the case of busy waiting, binding to an isolated core really helps reduce the higher latency gaps.

These tests were run at the same time.  The only difference is that the "bound" thread was bound to an "isolcpus" CPU where the other CPU for that core was isolated as well. i.e. the whole core was isolated.

Fairly busy - 50%

In this case, the thread alternated between sampling for 1 milli-second and sleeping for 1 milli-second
The unbound 50% busy thread had much lower delays of 2 micro-seconds, but significantly more of longer delays.

Slightly busy - 10%

In this test, the sampler runs for 0.111 millis-seconds and sleeps for 1 milli-second.  Even in this case, binding to an isolated CPU makes a difference.



Binding but not isolating - 10%

In this case, the bound thread was not isolated.  It was bound to a CPU where the core wasn't free and it was not isolated either.  It would appear that binding alone makes little difference compared to unbound in this test.

Comparing bound and isolated threads

Something I have seen before, but I find a little strange is that if you give up the CPU your thread performs badly when it wakes up.  Previously I had put this down the caches not being warmed up, but the code does very little access to memory and the code is very short so it is still possible but unlikely.  The peak at 20 micro-seconds at one million per hour could be due to a delay which happens on every forth wake up. It is about 90,000 clock cycles which seems like a lot for a cache miss.

Comparing unbound threads

In this chart, it suggest it really helps to be greedy of the CPU even if you are not bound.  Busy threads get interrupted less.  It is hard to say that 50% busy is better than 10% busy.  It might be, but a longer test would be needed (I would say it's within the margin of error)

Conclusion

Using thread affinity, without isolating the CPU doesn't appear to help much on this system.  I suspect this is true of other versions of Linux and even Windows.

Where affinity and isolation helps, it may still make sense to busy wait as it appears the scheduler will interrupt the thread less often if you do.


Comments

  1. Thierry Abaléa5 July 2013 at 21:26

    Taking also into consideration the results of your previous post "How much difference can thread affinity make" (http://vanillajava.blogspot.fr/2012/02/how-much-difference-can-thread-affinity.html), the conclusion is that: thread affinity without cpu isolation improved latency and thread affinity with cpu isolation improved jitter and throughput. Am I right ?

    ReplyDelete
  2. Peter Lawrey5 July 2013 at 22:07

    I wrote the post in response to the view that pinning CPUs is for the weak or shouldn't be needed.

    @Thierry I suspect the real impact will depend on what your application is doing. In the test above, there is a very big difference in using an isolated CPU but not such much just bonding to a CPU which was not isolated. The test is not realistic at all. What the test does suggest is that the OS looks at what the program appears to be doing and behaves differently. The isolated/bonded threads which don't busy wait get more delays in the 14-40 micro-second range.

    In truth you need to test this for your application, your OS on your hardware. The point of the post is to show that using an isolated CPU can make a huge difference, at least in some cases, and it is well worth considering if sub-milli-second latency is important to you.

    ReplyDelete
  3. Thierry Abaléa5 July 2013 at 22:46

    Thanks for the clarification.

    It appears that even applications that don't required sub-milli-second latency can take benefit of CPU isolation. They can scale at a lower cost. Quora use this technique for their web servers http://engineering.quora.com/Web-Server-Architecture-at-Quora

    ReplyDelete

Post a Comment

Popular posts from this blog

Java is Very Fast, If You Don’t Create Many Objects

By
Image
  You still have to watch how many objects you create. This article looks at a benchmark passing events over TCP/IP at 4 billion events per minute using the net.openhft.chronicle.wire.channel package in Chronicle Wire and why we still avoid object allocations..  One of the key optimisations is creating almost no garbage. Allocation is a very cheap operation and collection of very short-lived objects is also very cheap. Does this really make a difference? What difference does one small object per event (44 bytes) make to the performance in a throughput test where GC pauses are amortised? While allocation is as efficient as possible, it doesn’t avoid the memory pressure on the L1/L2 caches of your CPUs and when many cores are busy, they are contending for memory in the shared L3 cache.  Results Benchmark on a Ryzen 5950X with Ubuntu 22.10. JVM Vendor, Version No objects Throughput, Average Latency* One object per event Throughput, Average Latency* Azul Zulu 1.8.0_322 60.6 M event/s, 528
Read more

System wide unique nanosecond timestamps

By
A Unique Identifier can be very useful for tracing. Those ids are even more useful when they contain a high-resolution timestamp.  Not only do they record the time of an event, but if unique can help trace events as they pass through the system. Such unique timestamps however can be expensive depending on how they are implemented.   This post explores a lightweight means of producing a unique, monotonically increasing system-wide nano-second resolution timestamp available in our open-source library. Uses for Unique Identifiers Unique identifiers can be useful to associate with a piece of information so that information can be referred to later unambiguously. This could be an event, a request, an order id, or a customer id. They can naturally be used as a primary key in a database or key/value store to retrieve that information later. One of the challenges of generating these identifiers is avoiding creating duplicates while not having an increasing cost.  You could record every identi
Read more

Comparing Approaches to Durability in Low Latency Messaging Queues

By
Image
  A significant feature of Chronicle Queue Enterprise is support for TCP replication across multiple servers to ensure the high availability of application infrastructure. I generally believe that replicating data to a secondary system is faster than syncing to disk, assuming the round trip network delay wasn’t high due to quality networks and co-located redundant servers. This is the first time I have benchmarked it with a realistic example. Little’s Law and Why Latency Matters In many cases, the assumption is that the latency won't be a problem as long as throughput is high enough. However, latency is often a key factor in why the throughput isn’t high enough. Little’s law states, “ the long-term average number  L  of customers in a  stationary  system is equal to the long-term average effective arrival rate  λ  multiplied by the average time  W  that a customer spends in the system”. In computer terminology, the level of concurrency or parallelism a system has to support must be
Read more