Community discussions

MikroTik App
 
SystemErrorMessage
Member
Member
Topic Author
Posts: 383
Joined: Sat Dec 22, 2012 9:04 pm

CRS switching performance and method

Sun Dec 14, 2014 4:42 pm

Im thinking about getting a CRS for my cluster because my older gigabit managed switch cannot have more than 4 combined ports (1 combined port per logical section). I mainly want to know the switching delay for a packet on the CRS (just LAN transfers) and if it can use both cut through or store and forward switching methods and if it is possible to combine ports using the switch chip because i expect to use the full wire.

How much can the CRS do at wirespeed that a managed switch can do? The specs are also confusing because different CRS models appear to have different CPU speeds and ram.
 
jarda
Forum Guru
Forum Guru
Posts: 7756
Joined: Mon Oct 22, 2012 4:46 pm

Re: CRS switching performance and method

Sun Dec 14, 2014 7:09 pm

Unfortunately I don't know the answer but I am sure that cpu and ram has nothing to do with switching as it is done in switch chip and not in cpu.
 
SystemErrorMessage
Member
Member
Topic Author
Posts: 383
Joined: Sat Dec 22, 2012 9:04 pm

Re: CRS switching performance and method

Sun Dec 14, 2014 7:32 pm

Yes i know its unrelated but personally i feel bad if a model line that is supposed to have the same hardware but different interfaces have different things like different ram amounts, etc. The most important thing really i want to know is how good is CRS as a switch if it has very low latency switching and switching method. To me this is very important really in the way that i set things up for performance. My current switch has 7microseconds of latency on average per packet but uses store and forward which means i have to use the default 1500byte packets. For managed switches switching latency, non-blocking performance and switching methods are mentioned. I think mikrotik needs to look into getting their switches benchmarked. I would prefer to use larger packets to minimise CPU for transfers so that CPU can focus on more important things like computation and services. Also because many of my devices utilise more than 1 port i need port trunking on wirespeed for many devices which i think routerOS can do on the CRS at the switch level unlike my older switch that can only do LAGG on 1 device per switch segment. That means 4 devices max can have combined ports for additional bandwidth.

This isnt for a rendering cluster, it is for a computational cluster where latency is very important and doesnt need to have 10G links. A single ping may show a very low latency but if you tried to ping large sizes or lots of data you will find that latency is quite an issue even in LAN.
 
gsauthof
just joined
Posts: 9
Joined: Fri May 28, 2021 8:00 pm
Contact:

Re: CRS switching performance and method

Sun May 30, 2021 2:38 pm

Nowadays, the Mikrotik CRS switch product pages do include performance reports. See for example https://mikrotik.com/product/crs326_24g ... estresults

Those reports don't include latency figures, though. Just non-blocking throughput/capacity.

Latency might also vary much between different ports/port-groups (cf. the block diagram, e.g. https://i.mt.lv/cdn/product_files/CRS32 ... 200427.png) and on the traffic currently switches on other ports.

But still, it would be helpful if Mikrotik would also publish some latency test results.

It's probably safe to say that most switches are configured for store-and-forward, by default. And that cut-through switching is a premium feature. Thus, if a switch supports it, it's usually explicitly pointed out by the vendor in the technical specification. So since Mikrotik doesn't document store-forward/cut-through configuration settings anywhere (as of 2021), it's likely that no available Mikrotik device currently supports it.

You could hunt down the datasheet for the switch's packet processor CPU (e.g. Marvell 98DX3236 for CRS326) to see if it even could support cut-through switching, but even if it did, Mikrotik might still decide to not expose it for various reasons.

It probably doesn't hurt to e-mail the Mikrotik support to inquire about any future-plans for offering cut-through switching.

Regarding 1 GBit vs. 10 GBit latency - even if you don't need the 10 GBit line-rate, 10 GBit ethernet still offers better latency than 1 GBit ethernet, because both use single-lane transmission and higher speed reduces (de-)serialization time. Similar argument can be made for 25 GBit vs 10 GBit with respect to host-to-host latency.

Who is online

Users browsing this forum: No registered users and 18 guests