Juniper MX Sizing

Brad Fleming bdflemin at gmail.com
Fri Dec 5 21:10:40 UTC 2014


We haven’t received the MX480 gear yet (POs just went in about a week ago). But we tested MX960s with the same RE-S-1800x4 w/ 16GB RAM RIB+FIB convergence time was roughly 45sec. We never worried about getting a super accurate time for the MX960 because even an “eye test” showed it was fast enough for our application and we were much more concerned with other parts of the box. Also, we had inline-flow reporting configured on the MX960. Actually, the MX960’s had a full, production-ready config while the MX104 was tested with a stripped down after we discovered the slow convergence.

Once we get some MX480s on the bench I’ll report back.


> On Dec 5, 2014, at 2:35 PM, Shawn Hsiao <phsiao at tripadvisor.com> wrote:
> 
> 
> MX480 is also not instantaneous, so the same problem applies.   Brad, do you have the number for MX480 for comparison?
> 
> What we decided was, given both models suffer the same problems, just different duration, we decided to mitigate the problem and not spending the money.
> 
> Thanks.
> 
> 
> 
> On Dec 5, 2014, at 3:01 PM, Brad Fleming <bdflemin at gmail.com> wrote:
> 
>> Then you should look for something other then the MX104.
>> 
>> In our testing an MX104 running Junos 13.3R4 with a single, full feed took about 4min 25sec to (1) converge the RIB from a router sitting 0.5ms RTT away and (2) update the FIB with all entries. This performance was observed with single RE and dual RE and without any excess services running. If we added inline-flow sampling to the device full convergence took closer to 5min 45sec in our lab. Efforts to bring the convergence time down (without filtering ingress advertisements) with the assistance of JTAC proved unsuccessful.
>> 
>> We decided to “bite the bullet” and procure MX480s instead but obviously that’s not possible for everyone. If the MX480 is out of the question a Brocade CER Premium is an option. We have 3 in production and see very attractive convergence times; however, they have a more limited feature set and you’ll want to understand how their FIB memory scales. Apologies, I don’t know the Cisco equivalent from the ASR line these days but I’m sure others on the list could help out.
>> 
>> 
>>> On Dec 5, 2014, at 11:45 AM, Graham Johnston <johnstong at westmancom.com> wrote:
>>> 
>>> Shawn,
>>> 
>>> It's more about FIB than RIB as I am concerned about the time it takes until MPCs have updated route information after large scale changes in routes learned via BGP.
>>> 
>>> Graham Johnston
>>> Network Planner
>>> Westman Communications Group
>>> 204.717.2829
>>> johnstong at westmancom.com
>>> think green; don't print this email.
>>> 
>>> -----Original Message-----
>>> From: Shawn Hsiao [mailto:phsiao at tripadvisor.com] 
>>> Sent: Friday, December 05, 2014 11:30 AM
>>> To: Graham Johnston
>>> Cc: nanog at nanog.org
>>> Subject: Re: Juniper MX Sizing
>>> 
>>> 
>>> Is your sizing concern just for the RIB, or also for FIB to sync up?   The latter was a problem for us, but not the former.   We also have inline-jflow turned on and that is still a work-in-progress in terms of impacting performance.
>>> 
>>> We are using MX104 for similar purposes for many months now, and with some tweaks in our procedures and configurations we found it to be acceptable.    MX104 may not be able to process routes as fast as MX480, but MX480 is also not instantaneous either so similar risks exist.
>>> 
>>> 
>>> 
>>> On Dec 5, 2014, at 11:59 AM, Graham Johnston <johnstong at westmancom.com> wrote:
>>> 
>>>> I am wondering if anyone can provide their real world experience about sizing Juniper MX routers as it relates to BGP.  I am needing a device that has a mix of layer 2 and 3 features, including MPLS, that will have a very low port count requirement that will primarily be used at a remote POP site to connect to the local IX as well as one or two full route transit providers.  The MX104 has what I need from a physical standpoint and a data plane standpoint, as well as power consumption figures.  My only concern is whether the REs have enough horsepower to churn through the convergence calculations at a rate that operators in this situation would find acceptable.  I realize that 'acceptable' is a moving target so I would happily accept feedback from people using them as to how long it takes and their happiness with the product.
>>>> 
>>>> For those of you that deem the MX104 unacceptable in this kind of role and moved up to the MX240, what RE did you elect to use?
>>>> 
>>>> Thanks,
>>>> Graham Johnston
>>>> Network Planner
>>>> Westman Communications Group
>>>> 204.717.2829
>>>> johnstong at westmancom.com<mailto:johnstong at westmancom.com>
>>>> P think green; don't print this email.
>>>> 
>>> 
>> 
> 




More information about the NANOG mailing list