Chronicle Map and Yahoo Cloud Service Benchmark

Overview

Yahoo Cloud Service Benchmark is a reasonably widely used benchmarking tool for testing key value stores for a significant number of key e.g 100 million, and a modest number of clients i.e. served from one machine.

In this article I look at how a test of 100 million * 1 KB key/values performed using Chronicle Map on a single machine with 128 GB memory, dual Intel E5-2650 v2 @ 2.60GHz, and six Samsung 840 EVO SSDs.

The 1 KB value consists of ten fields of 100 byte Strings.  For a more optimal solution, primitive numbers would be a better choice. While the SSDs helped, the peak transfer rate was 700 MB/s which could be supported by two SATA SSD drives.

These benchmarks were performed using the latest version at the time of the report, Chronicle Map 2.0.6a-SNAPSHOT.

Micro-second world.

Something which confounds me when reading benchmarks about key-value stores is that they start with the premise that performance is really important.  IMHO, about 90% of the time, performance is not the most important feature, provided you have sufficient performance.
These benchmark reports then continue to report times in milli-seconds, not micro-seconds and throughputs in the tens of thousands instead of the hundreds of thousands or millions.  If performance really was that important, they would have built their products around performance, instead of the useful features they do support, like multi-key transactionality, quorum updates and other features Chronicle Map doesn't support, for performance reasons.

So how would a key-store built for performance look with YCSB?

Throughput measures

The "50/50" tests 50% random reads and 50% random writes, the "95/5" tests 95% reads to 5% writes. It is expected that writes will be more expensive, and a higher percentage of reads results in higher throughputs.

Threads 50/50 read/update 95/5 read/update
1 122 K/s 262 K/s
2 235 K/s 496 K/s
4 339 K/s 910 K/s
8 565 K/s 1.010 M/s
15 973 K/s 1.445 M/s
30 816 K/s 1.787 M/s

Latencies

The following latencies are in micro-seconds, not milli-seconds.

Threads: 8 50/50 read 95/5 read 50/50 update 95/5 update
average 5.7 µs 4.9 µs 13 µs 12.9 µs
95th 15 µs 13 µs 27 µs 25 µs
99th 25 µs 30 µs 44 µs 47 µs
worst 52 ms 52 ms 52 ms 52 ms

Note: the benchmark is not designed to be GC free and creates some garbage.  This is not particularly high and the benchmark itself uses only about 1/4 of CPU according to flight simulator, however it does impact the worst latencies.

Conclusion

Make sure the key-value store has the features you need, but if performance is critical, look for a solution designed for performance as this can be 100x faster than full featured products.

Other high performance examples

Aerospike benchmark - Single server benchmark with over 1 M TPS, sub-micro-second latencies. Uses smaller 100 byte records.
NuoDB benchmark - Supports transactions across a quorum. 24 nodes for 1 M TPS.
Oracle NoSQL benchmark - A couple of years old, uses a lot of threads, otherwise a good result.
VoltDB benchmark - Not tested to 1 M TPS, but promising. Latencies around 1-2 ms, report has 99th percentile latencies which others don't include.

Room for improvement

MongoDB driver benchmark - Has 1000s of micro-seconds instead of milli-seconds.
Cassandra, HBase, Redis - Shows you can get 1 million TPS if you use enough servers, 288 nodes for 1 M TPS.
Report including Elasticsearch - Report includes runtime in a "resource Austere Environment"
Hyperdex - Cover throughput only.
WhiteDB - Reports latencies in micro-seconds for 170 K records, and modest throughputs.
Benchmark including Aerospace - Reports 

Footnote

Using smaller values helps, and we suggest trying to make values closer to 100 bytes.  This is the result of the 95/5 workload B, using 10x10 byte fields, and 50 M entries as the Aerospike benchmark does. (30 clients)

[OVERALL], RunTime(ms), 60,669
[OVERALL], Throughput(ops/sec), 3,296,576
[READ], Operations, 190002671
[READ], AverageLatency(us), 4.81
[READ], MinLatency(us), 0
[READ], MaxLatency(us), 74864
[READ], 95thPercentileLatency(ms), 0.009
[READ], 99thPercentileLatency(ms), 0.014
[READ], Return=0, 115841209
[READ], Return=1, 74161462
[UPDATE], Operations, 9997309
[UPDATE], AverageLatency(us), 12.23
[UPDATE], MinLatency(us), 1
[UPDATE], MaxLatency(us), 75015
[UPDATE], 95thPercentileLatency(ms), 0.017
[UPDATE], 99thPercentileLatency(ms), 0.028
[UPDATE], Return=0, 9997309






Comments

Post a Comment

Popular posts from this blog

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

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 ...
Read more

System wide unique nanosecond timestamps

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 ...
Read more

What does Chronicle Software do?

Image
Overview Chronicle Software is about simplifying fast data.  It is a suite of libraries to make it easier to write, monitor and tune data processing systems where performance and scalability are concerned. But its free, how do you make money, through support? We offer premium support . However, we often hear from users for the first time about a year after they have it in production.  Most users find they can support the software themselves.  It is only after a year or so, they have questions or concerns about where to take the product next.  In many cases, it is to a problem we have already solved and they just need to upgrade their software or use a new module we have added. As we don't know who is using most of our software, we can't advise them on how best to use our software and what updates or enhancements they would benefit from. We need users to contact us and ask questions.  We have a free forum , and we respond to 50% of questions in 2 hours...
Read more