[
https://issues.apache.org/jira/browse/CASSANDRA-15299?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17220921#comment-17220921
]
Sam Tunnicliffe commented on CASSANDRA-15299:
---------------------------------------------
Here are some numbers from running tlp-stress against 4.0-beta2 vs the 15299
branch {{70923bae4}} with protocol v4 & v5.
The dataset was sized to be mostly in-memory and reading/writing at {{CL.ONE}}.
Each workload ran for 10 minutes, with a 1m warmup, on a 3 node cluster.
Both latency and throughput are pretty much on a par, with v5 showing a bit of
an improvement where the workload is amenable to compression.
Memory usage and GC were pretty much the same too. If anything, during the v5
runs the servers spent less time in GC, but this was so close as not be
significant.
There are definitely some places we can improve the performance of v5, and I
don't think anything here indicates a serious regression in either v4 or v5
performance. Given the additional integrity checks in v5, I think not
regressing meets the perf bar here, at least in the first instance.
h4. 100% read workload, reads a single row at a time.
{code}
Count Latency (p99) 1min (req/s)
4.0-beta2 V4 135449878 72.43 224222.86
15299 V4 138424112 68.3 229765.45
15299 V5 137618437 70.35 231602.36
4.0-beta2 V4 LZ4 103348953 66.68 173437.38
15299 V4 LZ4 105114560 68.83 176192.36
15299 V5 LZ4 131833462 70.19 222092.99
{code}
h4. Mixed r/w workload (50/50). K/V with blobs upto 65k
{code}
Count Latency (p99) 1min (req/s) | Count
Latency (p99) 1min (req/s)
4.0-beta2 V4 34455009 76.59 56557.78 | 34464217
74.85 56546.40
15299 V4 33368171 74.90 54859.94 | 33361940
67.77 54848.57
15299 V5 32991815 76.00 54780.74 | 33001153
76.72 54856.99
4.0-beta2 V4 LZ4 32152220 83.41 53306.03 | 32147206
83.22 53334.00
15299 V4 LZ4 31158895 71.01 51106.60 | 31153453
72.75 51087.01
15299 V5 LZ4 32634296 75.73 54370.71 | 32644765
76.70 54396.15
{code}
h4. Mixed r/w/d workload (60/30/10). Wide rows with values up to 200b and
slicing on both the inserts and deletions.
{code}
Count Latency (p99) 1min (req/s) | Count
Latency (p99) 1min (req/s) | Count Latency (p99) 1min (req/s)
4.0-beta2 V4 18725688 197.47 25377.16 | 9357971
193.86 12687.94 | 3117394 178.44 4221.90
15299 V4 17975636 185.88 24160.78 | 8986125
184.97 12087.13 | 2995562 179.99 4023.88
15299 V5 18429252 192.91 25349.35 | 9223277
188.15 12678.46 | 3073312 184.24 4232.23
4.0-beta2 V4 LZ4 18407719 179.94 25160.16 | 9197664
178.25 12575.03 | 3068134 180.90 4195.38
15299 V4 LZ4 17678994 171.39 24073.09 | 8842952
196.06 12064.18 | 2947344 170.53 4026.37
15299 V5 LZ4 18274085 208.57 25127.02 | 9138491
163.23 12558.28 | 3045264 203.54 4188.88
{code}
> CASSANDRA-13304 follow-up: improve checksumming and compression in protocol
> v5-beta
> -----------------------------------------------------------------------------------
>
> Key: CASSANDRA-15299
> URL: https://issues.apache.org/jira/browse/CASSANDRA-15299
> Project: Cassandra
> Issue Type: Improvement
> Components: Messaging/Client
> Reporter: Aleksey Yeschenko
> Assignee: Sam Tunnicliffe
> Priority: Normal
> Labels: protocolv5
> Fix For: 4.0-alpha
>
>
> CASSANDRA-13304 made an important improvement to our native protocol: it
> introduced checksumming/CRC32 to request and response bodies. It’s an
> important step forward, but it doesn’t cover the entire stream. In
> particular, the message header is not covered by a checksum or a crc, which
> poses a correctness issue if, for example, {{streamId}} gets corrupted.
> Additionally, we aren’t quite using CRC32 correctly, in two ways:
> 1. We are calculating the CRC32 of the *decompressed* value instead of
> computing the CRC32 on the bytes written on the wire - losing the properties
> of the CRC32. In some cases, due to this sequencing, attempting to decompress
> a corrupt stream can cause a segfault by LZ4.
> 2. When using CRC32, the CRC32 value is written in the incorrect byte order,
> also losing some of the protections.
> See https://users.ece.cmu.edu/~koopman/pubs/KoopmanCRCWebinar9May2012.pdf for
> explanation for the two points above.
> Separately, there are some long-standing issues with the protocol - since
> *way* before CASSANDRA-13304. Importantly, both checksumming and compression
> operate on individual message bodies rather than frames of multiple complete
> messages. In reality, this has several important additional downsides. To
> name a couple:
> # For compression, we are getting poor compression ratios for smaller
> messages - when operating on tiny sequences of bytes. In reality, for most
> small requests and responses we are discarding the compressed value as it’d
> be smaller than the uncompressed one - incurring both redundant allocations
> and compressions.
> # For checksumming and CRC32 we pay a high overhead price for small messages.
> 4 bytes extra is *a lot* for an empty write response, for example.
> To address the correctness issue of {{streamId}} not being covered by the
> checksum/CRC32 and the inefficiency in compression and checksumming/CRC32, we
> should switch to a framing protocol with multiple messages in a single frame.
> I suggest we reuse the framing protocol recently implemented for internode
> messaging in CASSANDRA-15066 to the extent that its logic can be borrowed,
> and that we do it before native protocol v5 graduates from beta. See
> https://github.com/apache/cassandra/blob/trunk/src/java/org/apache/cassandra/net/FrameDecoderCrc.java
> and
> https://github.com/apache/cassandra/blob/trunk/src/java/org/apache/cassandra/net/FrameDecoderLZ4.java.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org
