Which Queue Can Use the Switch Buffer?

Start with packets waiting for an output

When packets need an output that cannot serve them immediately, buffering lets admitted traffic wait. The queue represents waiting traffic; packet memory supplies the finite storage backing it. A brochure's buffer-capacity figure identifies a storage quantity, but leaves open which traffic can occupy it together.

The useful buying question is therefore specific: can the queues carrying the intended simultaneous traffic obtain enough usable memory while their packets wait? A demonstration that one queue can grow when the others are quiet answers a different question from concurrent demand.

Keep storage and transmission service separate from the start. RFC 2309, an April 1998 informational document about Internet routers, distinguishes managing queue length from scheduling which packet to send next. That conceptual distinction is useful here without importing its router algorithms into an Ethernet switch.

References: [R1]

Separate physical space, reservations and admission

Read an offer through four quantities. Total physical space is the storage within the stated hardware boundary. A private reservation protects an assigned portion for its stated consumer. Shared-pool availability is the unoccupied space that eligible consumers can compete to use. An admission allowance is a policy boundary on accepting additional traffic. An allowance and available storage must both be considered.

NVIDIA's Cumulus Linux 5.14 QoS guide provides a named vocabulary: traffic pools associate switch priorities with memory allocations, while buffer-region properties distinguish reserved allocation from static and dynamic shared allocation. It documents ingress and egress regions separately. This vocabulary helps frame a supplier's explanation; it does not establish universal ASIC accounting or portable defaults.

A queue can be permitted to draw on shared storage without owning that storage in advance. Another eligible queue may already occupy it. Conversely, free memory somewhere in the device does not establish that the queue is allowed to use it: pool mappings and admission boundaries still matter. Have the supplier explain the relationship between logical accounting regions and physical storage before adding figures from different regions.

References: [N1][C1]

Work the 8 MiB case under simultaneous demand

Invent a device with exactly 8 MiB of packet memory: queue A has a 1 MiB private reservation, queue B has a separate 1 MiB private reservation, and both can use one 6 MiB shared pool. Stipulate that these are the only allocations and consumers. Each queue's policy permits up to 4 MiB of shared occupancy when space is available.

All quantities in this example are abstract admitted-memory requirements, already expressed in the same memory units. They are not application-payload sizes, packet-to-cell conversions, product specifications or recommended settings. MiB denotes 1,048,576 bytes. Compare resident requirements at one instant, without crediting departures during that snapshot.

First suppose A requires 5 MiB and B requires 1 MiB. A needs its 1 MiB reservation plus 4 MiB shared; B fits within its own reservation. Total resident demand is 6 MiB, including 4 MiB shared, so this demand fits the stipulated space and allowances. The shared pool still has 2 MiB free.

Now suppose A and B each require 5 MiB at the same instant. Each needs 4 MiB beyond its reservation. Shared demand becomes 4 + 4 = 8 MiB against 6 MiB available: a 2 MiB shortage. Counting the reservations gives the same result: 1 + 1 + 8 = 10 MiB demanded against 8 MiB physical space.

Neither queue individually exceeds its stipulated 4 MiB shared allowance, yet their combined requirements cannot coexist. If A already occupies 4 MiB shared, only 2 MiB shared remains for B while A's occupancy is unchanged. The arithmetic does not promise that B will be admitted to that remaining space; a real admission policy could restrict it further. Nor does the shortage identify particular dropped packets or require an equal split. It establishes that two individually permitted peaks are not a simultaneous storage guarantee.

Space becomes reusable as stored traffic leaves

The snapshot changes when packets receive service and their accounted storage is released. Shared memory freed by A can become available to eligible traffic from B, subject to the applicable policy. If A retains its backlog because service is delayed, its occupancy continues to constrain what can fit alongside it. Timing matters even when the headline memory total is unchanged.

Cumulus Linux 5.14 documents egress scheduling separately from buffer allocation: the scheduler governs transmission from the queues. A memory allowance therefore cannot be read as an assigned output rate.

Additional usable memory can hold more temporary backlog under otherwise unchanged assumptions. It cannot make a fixed-capacity output transmit faster. If traffic persistently arrives faster than that output serves it, finite extra storage cannot preserve an ever-growing backlog indefinitely. The linked switch-bandwidth article addresses populated uplink capacity; this article addresses how waiting traffic competes for storage before it leaves.

References: [N1][R1]

Identify the offered hardware's usable admission boundary

The simple example gives every byte one role. A real offer needs its own accounting explanation. Request the exact switch and ASIC or line-card variant, software release, memory units and pool mappings for the traffic being compared. Ask what a queued packet is charged in the hardware's accounting, rather than equating an application byte count with consumed packet memory.

Cisco's Nexus 9300-H white paper, updated January 30, 2024, illustrates why the implementation matters. Its CloudScale H2R/H1 discussion describes cell-based on-die storage and Dynamic Buffer Protection that can deny admission at a queue limit. It also distinguishes on-die buffering from the H2R's separately accounted deep-buffer system. These are named architecture examples, not rules to transfer to another platform.

Have the supplier identify protected reservations, applicable headroom, consumers sharing the pool, and queue, port or hardware-partition limits that can constrain access. Explain which quantities overlap in accounting and which are physically separate. Memory outside the relevant admission boundary cannot simply be added to the queue's usable space.

Retain the documented admission, discard and congestion-marking behavior for the intended traffic mode. Physical memory remaining in the device does not mean every arriving packet must be accepted. The linked RoCE article explains the separate pause, ECN and CNP mechanisms; their presence alone does not establish how much memory a queue can use.

References: [N1][C1]

Compare the offers against the same competing traffic

Multiplying the brochure total by the number of ports counts the same shared storage repeatedly. Dividing it equally by the port count makes a different unsupported assumption: that every port has an identical private allocation. Sharing can let an eligible queue use more than an equal division when competitors are quiet, while simultaneous demand can leave it with less.

For each offer, identify which queues can contend for the same pool and what resident demand matters together. Keep the traffic population, packet-accounting assumptions and output-service conditions consistent. A comparison of one isolated queue with many active queues would hide the contention exposed by the 8 MiB case.

An apparently larger pooled-memory device deserves a preference only when the supplier's evidence connects its usable admission boundary to the buyer's requirement. A larger total might be inaccessible to the relevant queues, divided across different partitions or subject to a tighter applicable limit. Equally, sharing may be useful when competing peaks do not overlap. Resolve those relationships before treating either the headline total or an isolated maximum as an equipment-selection verdict.

Ask the appointed network owner to assess the resulting explanation against the intended traffic and acceptance requirements. The example supplies a capacity contradiction, not a workload qualification or an instruction to change buffer settings.

Carry the selected connections into physical installation

Once the equipment and network owners select and approve the connection plan, Leviathan's structured cabling service can discuss the agreed fiber and copper installation, routing, dressing, labeling, physical checks and connection records. Buffer policy, switch configuration and network-performance acceptance remain with their appointed technical owners.

Use the service and project-inquiry links below to share the approved rack and port schedule, specified assemblies, materials supplier and delivery responsibilities, routes, access window and required physical handoff records. Those inputs turn the selected equipment into a defined cabling inquiry.

References: [L1]